14
vehicles, CVs with human drivers, individual CAVs, and manufacturing [365], [366],[367], [368]. In terms ofIoT data
cooperative CAVs. Traffic and driving behaviors are complex monitoring and anomaly detection, methods such as binary
to model and highly nonlinear and stochastic in nature. It is segmentation,dynamicprogramming,Bayesianmethods,non-
challenging to propose a generic model as traffic dynamics parametricmethods,andmodel-drivenmethodswereproposed
depend on the specific road geometries, whether it is signal- [369], [370]. For instance, Bayesian online change detection
ized,curvature,speedlimit,numberoflines,trafficmovements was used for the droplet video monitoring during inkjet print-
and approaches, etc. Even for traffic on the same road, the ing [371]. Workers play a significant role in manufacturing,
conditionscanchangedrasticallyaccordingtothetimeofday, warehouse and other systems. The consideration of human
events, or incidents. A small change in the road geometry can safetyandhumanperformanceandtheirimpacttothegeneral
greatly impact traffic patterns and reduce the accuracy of an system performance in engineering systems was investigated,
already calibrated and trained model using either traditional see for instance [372], [373], [374], [375], [376].
methods or AI-based approaches. Novel AGI techniques can ThedigitaltoolsinIoTandIndustry4.0enabledtheremote
significantly improve the accuracy of traffic modeling and andconnectedtrackingandcontrol,atanyplaceandplatform,
prediction and obtain generic models that can adapt to the by connecting the once isolated Operational Technology (OT)
conditions and variants mentioned above. Breakthroughs in networks to the rest of the enterprise through the Information
traffic modeling using AGI can also benefit the evaluation Technology (IT) network [377], [378]. While this allows for
of CAVs. The fidelity of simulation will be greatly improved a better manufacturing modeling, monitoring and control, it
to reduce the amount of real-world testing that is expensive, also poses profound questions in areas such as intellectual
time-consuming,andcausessafetyconcerns[346].Inaddition, property (IP) protection, data security and privacy. These
traffic models are usually highly nonlinear and complicated, issueshaveledtoproductrecalls,publicbacklash,lostrevenue
which makes it challenging to solve the associated control or for companies, and will potentially threaten public safety and
optimization problem in real-time. AGI can help identify an security [379], [380]. Specific techniques were proposed to
efficient model structure and be used to effectively search the address the above-mentioned challenge. Differential privacy
optimal control strategies for real-time implementation. This and privacy-preserving modeling are proposed to address the
is especially promising for autonomous driving tasks [347], privacy and security issue [381], [382], [383], [384]. For
which are traditionally divided into tasks, including localiza- instance,Go´mezetal.integratedtraditionaldifferentialprivacy
tion, perception, prediction, planning, and motion control. A withk-anonymity,sothatthepreserveddatacanbeaggregated
unified end-to-end AGI approach can greatly reduce the accu- and transmitted without risking the privacy [385]. Federated
mulative errors from each module and improve coordination Learning (FL) was used so that machines collaboratively train
amongdifferenttasks.SuchanAGI-basedautonomousdriving a model without sharing raw data [383], [384]. For instance,
model has already attracted much attention in academia and Liu et al. proposed a FL monitoring framework to enable
industry [348], [349], [350]. decentralizededgedevicestocollaborativelytrainananomaly
detection model [383]. Wang et al. applied FL to build a
universal anomaly detection model with each local model
G. Smart Manufacturing
trained by the deep reinforcement learning [384]. However,
The Industry 4.0 led by organizations such as GE and IBM it is reported FL suffers from data leakage [386], [387].
has allowed better production process description, communi- Though the machine learning and AI approaches have
cation,andcomputation,andthistrendisbeingacceleratedby shown useful in addressing a certain type of problem, it can
themassadoptionofIoTdevices,5Gcommunicationtechnol- be hard for them to work in general problems with varying
ogy,theadvancementofAIandAGItechnologies[351],[352], andheterogeneousenvironments,limitedannotateddata,mul-
[353], [354]. Accordingly, the Strategy for American Leader- timodal information source, etc. AGI poses opportunities to
ship in Advanced Manufacturing (AM) states that worldwide addressthesechallenges,andhasthepotentialtoprovidemore
competition in manufacturing has been dominated in recent comprehensive and robust solutions to further enhance the
decades by the maturation, commoditization, and widespread capabilities of existing AI based methods for manufacturing
application of computation in production equipment and lo- modeling,monitoringandcontrol.[388]surveyedtheAGIand
gistics. Industry4.0fields,andfoundthatthoughAGIstudieshashuge
With the IoT devices, various modeling, monitoring and potential, the gap between the AGI studies and the industry
control approaches have been proposed for manufactur- needs is high.
ing quality, performance and reliability improvements [355], We point out the potential of AGI in manufacturing here.
[356], [357], [358]. For instance, the product geometric accu- 1) Manufacturing system heterogeneity and model gen-
racy was investigated in [359] and different types of process eralization: The manufacturing processes and systems
properties, such as porosity, roughness, were investigated in can be complex with heterogeneous machines and op-
3D printing with various process sensing information from erations, and most AI methods may not be able to cope
IoT sensors [360], [361], [362]. The model sparsity and withtheheterogeneity.Thereisaneedtodevelopmodels
interpretability was considered during the sensor data qual- and methods that is generalizable to various scenarios.
ity modeling in semiconductor manufacturing [363], [364]. AGI has the potential to adapt to new scenarios, by
To address high dimensional sensor and video data, tensor transfer learning from one domain to another [389],
techniques and deep learning models are investigated in smart [390] or incorporating physics driven domain knowledge