Thesis (M.Sc)-Cairo University, 2024.

Bibliography: pages 109-112.

Fault Detection and diagnosis of faulty sensors readings in oil heating reactor
are conducted for early failure revelation and machine components preserving before
the damage. The processes of fault detection, diagnosis and correction especially
in oil heating reactor sensors are among the most crucial steps for reliable and
proper operation inside the reactor where they are considered as important tools
that enable the controller taking the best possible action to insure finally the output
quality. In this thesis, firstly fault detection of different types of sensors in oil
heating reactor using different types of faults with different levels is addressed.
Multiple approaches based on Neural Network (NN)s such as the classical Neural
Network (NN), Bidirectional Long Short Term Memory network (BiLSTM) based
on Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) are
suggested for this purpose. The suggested networks are trained and tested on real
dataset sequences taken from sensors array readings of real heating reactor in Egypt.
The performance comparison of the suggested networks is evaluated using different
metrics such as accuracy, precision, ... , etc . For faults detection and diagnosis
by the suggested networks, different types of faults are applied to sensors array
of oil heating reactor such as Bias, Stuck, Drift, precision degradation and Spike
fault. The suggested networks are simulated, trained and tested in this thesis using
MATLAB software 2021 and the advanced tool of "DeepNetworkDesigner”. From
the simulation results, the average accuracy of CNN, BiLSTM network and Classical
NN for detecting all faults types at all their different levels and to all tested sensors
are 98.6%, 82.1% and 88.2% respectively. The mentioned results prove that CNN
outperforms the other comparative networks with highest classification accuracy in
detecting all faulty sensors with all different types of faults and at all their different
levels. Drift fault was the easiest type to be detected by any one of the suggested
networks where the average accuracy of classical NN, BiLSTM network and CNN in
detecting all faulty sensors by this type of fault over its all levels are 98.8%, 96.6
and 100% respectively. and with average accuracy of all networks reached to 98.5On
the other hand, Spike fault was the hardest type to be detected by any one of the
suggested networks where the average accuracy of classical NN, BiLSTM network
and CNN in detecting all faulty sensors by this type of fault over its all levels are
52%, 79% and 96.6% respectively and with average accuracy of all networks reached
to 76.1%. The second part of this thesis is concerning by the fault type diagnosis.
After detecting which sensor are Faulty, the process of diagnosis is implemented
on this faulty sensor to determine the type of fault on it. From the findings of
diagnosis outputs , it is found that the accuracy of CNN for faults diagnosis to all
tested sensors was directly proportional to the fault level and ranged in average from
almost 73.9% at 1% fault level to 99.8% at 90% fault level.
CNN average for diagnosing Precision Degradation fault at its all levels and over all
tested sensors was the largest value and reached to 97.1% while was the lowest for
diagnosing Bias fault and reached to 85%. The simulation results prove that CNN
has the superiority in faults types diagnosis with average classification accuracy
reached to 91.6% calculated to all the faulty tested sensors and at all faults levels.

يتم اكتشاف القراءات الخاطئة لحساسات الحرارة المعطوبة بداخل مفاعل التسخين للزيت الخام بإستخدام تقنيات الذكاء الإصطناعي كتقنية الشبكة العصبية ذات الطبقات المتصلة كليا وتقنية الشبكة العصبية المتكررة ثنائية الإتجاة ذات الذاكرة القصيرة والطويلة المدي وإيضا الشبكة العصبية الإلتفافية.

Issues also as CD.

Text in English and abstract in Arabic & English.