According to the Industry 4.0 vision, big data management is among the new challenges for the factory of the future. While many approaches have been developed to investigate data analysis, data visualisation, data collection and management, the impact of big data exploration is still under-estimated. In this paper, we propose an approach for big data exploration in a dynamic context of interconnected systems, such as the Industry 4.0 domain. The approach relies on three main pillars: (i) a multi-dimensional model, that is suited for supporting the iterative and multi-step exploration of big data; (ii) novel data summarisation techniques, based on clustering; (iii) a model of relevance, aimed to focus the attention on relevant data only. Keyword: big data exploration, data summarisation, data relevance, Industry 4.0, smart manufacturing
Big data management challenges raised from the abundance of real time data
in Cyber-Physical Systems (CPS), enabled by the widespread di usion of IoT
technologies [
In this paper, we discuss the ingredients to enable exploration of real time data in a dynamic context of interconnected systems, where large amounts of data must be incrementally collected, organized and analysed on-the- y. Such
Spindle #1 Tool
Working stations (raw material is positioned here) Rotating drum
Engine that
moves
spindle on X,
Y andZ axes
ingredients have been already described in [
The paper is organised as follows: in Section 2 we will motivate the approach with the help of a smart factory case study and we highlight di erences with respect to related work; in Section 3 we will provide a general overview of the approach; the architecture of a framework built on top of the proposed techniques and models will be presented in Section 4 and preliminary experiments in the case study will be described in Section 5; nally, Section 6 closes the paper.
Let's consider an Original Equipment Manufacturer (OEM) producing multispindle machines. As shown in Figure 1, spindles work independently each other on the raw material, that is positioned on a rotating drum.
Each spindle is mounted on a unit moved by an electrical engine to perform X, Y and Z movements. The spindle rotation is impressed by an electrical engine and its rotation speed is controlled by the machine control. Spindles use di erent tools (that are selected according to the instructions speci ed within the Part Program) in order to complete di erent steps in the manufacturing cycle. For each spindle, the velocity of the three axes (X, Y and Z), the electrical current absorbed by each engine, the value of rpm for the spindle and the percentage of power absorbed by the spindle engine (charge coe cient) are measured. Hereafter, with term features we will refer to the measured quantities. The aim is to monitor the axle hardening of each spindle and the tool wear. Axle hardening is monitored by observing changes in the values of energy consumption (spindle engine charge coe cient) for similar rpm. By detecting di erences in energy consumption while using di erent tools, spindle hardening can be identi ed as the possible anomaly during the manufacturing operations. If the increase in energy consumption is related only to the usage of a speci c tool, this has been recognised as a symptom of tool wear. We collected real data from three machines, each one equipped with three spindles and di erent tools. On each spindle, we monitored the eight features listed above. 2.2
In the considered motivating scenario, experience of human operators still plays
a fundamental role. Data exploration must be performed in near real time and
during incremental data collection. In [
Data summarisation. The characteristics of big data, namely, volume,
velocity and variety, pose non trivial issues for data collection and organisation. High
volume calls for techniques and tools to provide a compact view over the large
amount of collected data. Furthermore, we deal with data streams, in which data
is collected in a fast and incremental way. We address these issues by applying
incremental clustering techniques. Clusters o er a two-fold advantage: (a) they
give an overall view over a set of measures, using a reduced amount of
information; (b) they allow to depict the behaviour of the system better than single
records, that might be a ected by noise and false outliers, in order to observe
a given physical phenomenon. Clustering is performed in two steps: (i) in the
rst one, a variant of Clustream algorithm [
Clustering-based data relevance evaluation. Relevance-based techniques
are used to focus exploration on relevant clusters only. In literature, data
relevance is de ned as the distance from an expected status. In our case, the
expected status corresponds to the normal working conditions of monitored
Cyber Physical Systems. The expected status can be tagged by human operators
while observing the monitored system. Let's denote with S^C = fC^1; C^2; : : : ; C^ng
the clusters set identi ed during normal working conditions, and with SC =
fC1; C2; : : : ; Cmg the current clusters set, where n and m do not necessarily
coincide. Relevant data are recognised when SC di ers from S^C. Therefore, the
proposed relevance techniques are based on clusters set distance between SC and
S^C, denoted with (SC; S^C), and enable to detect clusters movements, clusters
contraction/expansion, changes in the number of clusters. We refer to [
Multi-dimensional model for data exploration. The multi-dimensional model is shown in Figure 2, using the hypercube representation, where axes represent exploration dimensions and nodes represent clusters sets computed on measures. Multi-dimensional modeling may help to organise clustered data according to di erent perspectives, compliant with operators' requirements. Among the dimensions, we always consider feature spaces, time and operational parameters. With feature spaces we refer to monitored phenomena (e.g., axle hardening, tool wear) observed by measuring a set of features. Time is of paramount importance, since the data summarisation procedure and data relevance techniques are strictly related to the time dimension. Operational parameters represent domainspeci c settings or variables: for example, the working mode (G0, fast movement clusters set of measures asset hierarchy multi-spindle machine spindle1 spindle2
spindle3 % absorbed power time % absorbed power relevant data (e.g., different behaviour) rpm spindle2 (b) exploration across spindles % absorbed power ti
rpm (a) exploration over time (wco)rekxipnglormatoiodne over % absorbed power ti+tj relevant data (e.g., anomalous rpm behaviour) rpm spindle1 of the spindle to catch the tool, or G1, slow movement of the spindle during the manufacturing), the tool used during manufacturing, the monitored system, the manufactured product. Dimensions may present hierarchies: for example, tools can be aggregated into tool types, monitored system follows the asset hierarchy (enterprise, shop oor, machine, components).
Exploration scenarios. Data relevance is used to identify dimensions over which summarised data changed, thus denoting an unexpected behaviour of the monitored system. Nevertheless, data relevance evaluation performed over all available dimensions may bring to high computational requirements. Therefore, exploration scenarios are proposed to constrain relevance evaluation over a subset of available dimensions, to prevent useless comparisons. In Figure 2, some examples of exploration scenarios are highlighted. Exploration for anomaly detection (a) detects anomalies by observing if collected data of a speci c spindle changes and gets closer/overtakes physical limits of breakage. This exploration scenario constrains exploration over the time dimension and within the same monitored system. Exploration for performance comparison (b) focuses comparison between spindles working in the same conditions. Other exploration scenarios are de ned to avoid erroneous comparisons. For example, let's consider the exploration over working mode (c) as shown in Figure 2: a comparison of system behaviour across G0 and G1 modes does not make sense. Modeling exploration scenarios, in order to abstract from their implementation, will be further investigated in future work. 4
The approach described in this paper has been implemented within the IDEAaS (Interactive Data Exploration As a Service) framework, whose architecture is Sensor and
Control Data Data Acquisition Service
CPS Notification Service Anomaly Detection Service
Monitoring Service Data Control
Module
IDEAaS Core Modules Data Summarisation
Module
Data Relevance Evaluation Module
IDEAaS front-end Services
IDEAaS Cloud-based Services ModelMetadata
Col ected Summarised
Data Data shown in Figure 3. In the IDEAaS modular architecture, modules are distinguished in IDEAaS Core Modules (that include Data Control, Data Summarisation and Data Relevance Evaluation) and IDEAaS front-end Services, exposed both as web services and as standalone modules. Among these services, Data Acquisition Service is in charge of storing data collected from CPS. As shown in Figure 3, data coming from the physical system, properly collected through sensors and IoT technologies, is sent to the Data Acquisition Service to be processed. This service operates in order to minimise time spent for data acquisition. Specifically, data collected is rst saved as JSON documents in a MongoDB NoSQL database (Collected Data), minimising any other operation, namely data control and cleaning and data clustering that are performed in parallel. Dimensions of the multi-dimensional model are stored as metadata into the Model-MetaData relational database. On collected data, summarisation and relevance evaluation techniques are applied. Resulting clusters sets are stored as JSON documents within Summarised Data database (also using MongoDB technology). The CPS Noti cation Service is in charge of notifying to target operators data of interest according to the relevance evaluation techniques. Data relevance can be used to serve di erent purposes, ranging from anomaly detection to performance monitoring of CPS, for which speci c services will be designed in the future. On top of CPS Noti cation Service, a Data Exploration GUI will be developed as well, to allow operators to interact with the system and explore collected data. The IDEAaS architecture is implemented in Java, on top of a Glass sh Server Open Source Edition 4. 5
We tested the e ciency and e ectiveness of our approach in providing summarised data for exploration purposes, given the acquisition rate of records in
Spindle 1 Spindle 2 Spindle 3 dercprmeas coinnescnureemragpsyteion dercprmeas coinnescnureemragpsyteion dercprmeas coinnescnureemragpsyteion / / / / / / / / / / 15% 16% 20% 20% / / 29% 29% 30% 30% / / / / 40% 40% 20% 20% / / / / 30% 30% / / / / 20% 20% / / (a)
Cluster distances
Time response clustering 400000 the considered case study. We collected 140 millions of records from the three machines. All records present a timestamp and have been collected every 200ms (5 records per second). We run experiments on an Intel Core i7-6700HQ, CPU 2.60 GHz, 4 cores, 8 logical cores, RAM 16GB. Collected records of measures have been saved within MongoDB as JSON documents grouped into collections. Documents present a simple structure, with at most one level of depth, and collections have been organised considering the time as main dimension, in order to speed up both data storage and data extraction for clustering, that is applied on slots of records on a time interval t. This enabled to storage all 140 millions of records in 1 hour and 14 minutes, with an acquisition rate of 31,531 records per second. We tested clustering and hypercube generation on real data considering average values on 2 and 3 features (Figure 4a). The worse response time corresponds to the case where we performed clustering and relevance evaluation when no previous syntheses have been generated. Also in that case, IDEAaS framework was able to process 15,600 records in 11.5 seconds (processing rate of 1,356 records per second). Through the tasks of syntheses generation and clustering, the processed set of records is reduced to 7,2% on average. As shown in Figure 4b, we observed a variation in distance between clusters sets at the cost of decreasing the processing time to 255 records per second, that is still acceptable. To test e ectiveness of the IDEAaS framework in detecting working anomalies, we introduced unexpected working states to simulate spindle hardering by increasing energy consumption and decreasing rpm on a subset of collected data, as shown in Figure 5a. Experimental results depicted in Figure 5b show the clustering distance between clusters sets calculated at the time t and those calculated in case of normal working conditions. The gure shows how the techniques of the IDEAaS framework allow to timely identify the unexpected situations induced in the system under observation. 6
In this paper, we discussed the ingredients to enable exploration of real time data incrementally collected, organized and analysed on-the- y. Our approach combines: (i) a multi-dimensional model, that is suited for supporting the iterative and multi-step nature of data exploration; (ii) e cient data summarisation techniques, based on clustering; (iii) a model of relevance, to focus the attention on relevant data only. Future e orts will be devoted to the parallelisation of clustering and relevance evaluation by relying on data organisation in the multidimensional model, to the development of data visualisation techniques and to the de nition of di erent data exploration scenarios.