Big Data Meet Cyber-Physical Systems: A Panoramic Survey

In plain words

Cyber-physical systems are networks of sensors, devices and machines that measure and act on the physical world, such as traffic, power grids, factories and hospitals. Together they produce far more data than a single computer can handle comfortably. This survey is a broad tour of what happens to that data: how it is collected, stored, moved across networks, processed in the cloud or near the sensors, analysed, protected from attackers and kept energy-efficient. It is a map of the field, not a new system, and it does not test anything itself.

Its one worked example is about emergency response. When the normal cell network is down, phones can relay messages to each other directly. The authors work out how long a message takes to reach everyone and how reliably it gets there, in a simple mathematical model.

TL;DR

• A panoramic survey of big data in cyber-physical systems (CPS), covering collection, storage, processing, analytics, security and green design §I.
• It is a literature review. It presents no experiments and no hyperparameters, and most numbers it shows come from the works it cites §IV.A §VII.A §VIII.A.2 §VIII.A.3.
• Its only formal model is a device-to-device public-safety analysis of delay and reliability §IX.F eq. 1 eq. 6.
• Useful as an index of which technique fits which CPS problem tab. 3 tab. 5 tab. 7; do not treat its claims as verified results.

Problem

CPS data arrive in volumes and at speeds that exceed conventional systems. The authors argue that earlier surveys covered parts of this (IoT, cloud, security or green computing) but not big data in CPS as one subject §I §II tab. 2. The survey also treats three problems as open: security that is affordable for resource-limited devices, the cost of privacy-preserving analysis, and meeting latency and reliability targets for critical services §X.

Method

This is a narrative survey, not a method. The paper gives no search protocol, database, time window or inclusion criteria, so how the literature was chosen is not stated §I §II. The structure follows the data pipeline:

• Sources: context-aware sensing, remote sensing and social or crowd sensing §III fig. 1.
• Storage and communications: edge caching, caching for Hadoop-type systems, and network designs for machine-type and device-to-device traffic §IV.
• Processing: single-cloud and multi-cloud workflows, clustering, NoSQL, and fog computing §V tab. 3 tab. 4.
• Analytics: data mining, real-time streams, cloud analytics, spatio-temporal analysis and their tools §VI.
• Security and privacy: secure storage, access control, anomaly detection, homomorphic encryption and k-anonymity §VII tab. 5.
• Green design: energy-efficient collection, computing and processing, and green applications §VIII tab. 6 tab. 7 tab. 8.
• Applications: smart grids, military, city, medical, disaster and public safety §IX fig. 6.
• Challenges and conclusion §X §XI.

Section cites use the numbering that the IEEE class (IEEEtran) gives the paper: a Roman numeral for each section, a letter for each subsection, and a number after the letter for each sub-subsection. Equations, tables and figures are numbered in order of appearance.

The one model is the public-safety case study §IX.F. Phones form a device-to-device network when the base station is unavailable. A command center sends messages, and each cluster head relays them to its members. The authors model the delay of multicast rounds and the link reliability. The key steps are below.

# Public-safety multicast: total delay D (one reading)  [§IX.F] [eq. 2] [eq. 3]
# Input:  rates[i][l] = R_il (bit/s); access_time[i] = t_i (s)
#         theta_d = delay threshold theta_D (s); source = first sender
#         (which node starts the multicast is not stated)
# Output: total delay D (s); whether the packet was dropped

def multicast_delay(rates, access_time, theta_d, source):
    # Nodes that hold the packet; only the source holds it at the start  [§IX.F]
    holders = {source}
    # Unvisited nodes: the L-1 nodes that still have to be reached  [§IX.F]
    unvisited = all_nodes - holders
    delay = 0.0
    while unvisited:
        # Sender: the holder whose weakest link to an unvisited node is
        # strongest  [eq. 2]
        sender = argmax(
            holders,
            key=lambda i: min(rates[i][l] for l in unvisited),
        )
        # Which unvisited nodes decode this round (delta_i = 1)  [§IX.F]
        # The paper gives no per-round draw rule; see Pitfalls
        decoded = draw_decoded(sender, unvisited)
        missed = unvisited - decoded
        # A round costs the access time of the slowest node that missed it  [eq. 3]
        delay = delay + max((access_time[i] for i in missed), default=0.0)
        holders = holders | decoded
        unvisited = missed
        # Packet is dropped once the accumulated delay exceeds theta_D  [§IX.F]
        if delay > theta_d:
            return delay, True
    return delay, False

Key equations, in the paper's notation:

N^\star_j = \{\, i \mid \delta_i = 1 \,\} eq. 1

k^\star_j = \arg\max_{i \in N^\star_j} \Big\{ \min_{l \notin N^\star_j} R_{i,l} \Big\} eq. 2

D \leftarrow D + \max_{i \in \bar N_j} t_i, \qquad \bar N_j = \{\, i \in \Pi_u \mid \delta_i = 0 \,\} eq. 3

f_D(r) = 2\pi\lambda r\, e^{-\lambda \pi r^2}, \qquad r \ge 0 eq. 4

The text gives the out-of-range probability as p_R = e^{-\pi\lambda R^2} §IX.F. It also states the channel-error term as

p_e = 1 - \tfrac{1}{2} + \tfrac{1}{2}\,\operatorname{erf}\!\left( \frac{\beta_{th} - \alpha \times 10\log(r)}{\sqrt{2}\,\sigma} \right) eq. 5

and combines the terms as

R_L = p_e + p_R + p_{drop}, \qquad p_{drop} = \mathcal{P}(D > \theta_D) eq. 6

Symbols: \delta_i is 1 when node i receives the packet in the round; \Pi_u and \Pi_v are the unvisited and visited sets; t_i is the minimum access time of node i; \lambda is the density of devices; R is the transmission range; \alpha is the path-loss exponent; \sigma is the shadow-fading parameter; \beta_{th} is the minimum threshold for delivery §IX.F.

Hyperparameters

The paper sets no values for any model parameter. It also states no simulation settings, because it runs no simulations. Every value below is therefore "not stated" in the paper.

Results

No experiments are reported by this paper. The table lists the quantitative claims it makes. Each one is quoted from the work it cites and was not measured here. The "Baseline" column is empty because the survey gives no comparison of its own.

Paper vs code

No official code.

The TeX source has no code links, and the only URLs are the BibTeX and IEEE template mirrors. The paper names no project page. A GitHub repository search for the title words and for "cyber-physical big data" returned one repository, RSRawat06/cyber-physical-systems. It is a Java Hadoop water-quality project with no link to the authors, so it was not used.

Pitfalls

• No method to reproduce. Most sections summarise cited works in a sentence or two. Anyone who implements a technique must read the original paper, not this survey §III §V §VI.
• Selection is not described. The paper calls itself the first panoramic survey on this topic §I, but the literature selection is not documented, so coverage cannot be checked §II.
• Dated. The paper was posted in October 2018 and does not cover later work §I. Treat its tables as a snapshot.
• Quoted numbers are not checked here. The results above come from cited works, and their datasets and conditions are not described in this survey. The 98% and 100% caching figures are for one operator's data §IV.A.
• Summary tables are inconsistent in places. The clustering table places the minimum-spanning-tree method under centroid-based clustering tab. 3, while the text describes that same method as hierarchical single linkage §V.C. Read the text, not the table grouping, for the method family.
• Sum of probabilities. Eq. 6 adds three probabilities. A sum can exceed 1, and a sum is not in general the probability that several failure modes are all avoided. The paper gives no justification for the sum eq. 6.
• p_drop has no formula. Eq. 6 defines p_drop as the probability that the total delay exceeds θ_D, but the paper does not say how to compute it eq. 6 §IX.F.
• Sender selection is circular. Eq. 1 defines the received set of a round from the outcomes in that round, and eq. 2 chooses the sender from that set. The paper does not say whether the sender is chosen before the round or from the holders of the packet. The pseudocode uses the second reading, which is an interpretation, not a statement from the paper §IX.F.
• Set and count notation is mixed. The text sets the visited set to 1 and the unvisited set to L-1 §IX.F. Treat Π_v and Π_u as sets, not counts.
• σ is ambiguous. The text calls σ a variance, while eq. 5 uses √2σ as if it were a standard deviation eq. 5 §IX.F.
• Log base is missing. Eq. 5 writes 10 log(r) without a base. The 10 is the usual dB factor for a log base 10, but the paper does not say so eq. 5.
• Distance model assumes still devices. The Rayleigh distance law in eq. 4 assumes users do not move, as in the cited placement model §IX.F eq. 4.
• Latency and rate figures are requirements. The 5-10 ms latency, 99.99% reliability and rate figures come from cited standards and studies. They are targets, not outputs of the model §IX.F.
• Figures credit other sources. Figures 3 and 6 are marked as adapted from other works fig. 3 fig. 6. This card does not use them and describes no figure content beyond what the text states.
• Costs the abstract hides. The paper says homomorphic encryption is expensive and impractical for large datasets §VII.B §X. It also says encryption makes deduplication harder, and cites a token-based scheme for it §VII.A.

Lineage

• Builds on: the device-to-device placement model cited as [6817780] for eq. 4 and the channel-error expression cited as [6093225] for eq. 5 §IX.F. Neither is in this wiki.
• Compared against: the earlier surveys listed in tab. 2, covering IoT, context-aware computing, mobile crowd-sensing, caching, cloud schedulers, big data analytics, security of MapReduce, fog computing, smart-grid security and green computing.
• Superseded by: not known from content already in this wiki.