Title: Development of Data driven Digital Twin for 3D Printers using Explainable AI

I. Statement of the problem

i. State the main problem you seek to address:
Additive manufacturing is emerging due to its advancement, such as rapid prototyping,
manufacturing flexibility, sustainability, and cost-effective technology. 3D printed
products still subjected to defects due to variation in properties and structure, which
leads to deterioration in the quality of printed products. Along with this due to internal
system faults and mechanical degradation of 3D printer components also leads to defects
in end products.
ii. Who has this problem, where does it occur?
3D printed products are widely used in the sectors of manufacturing, medical, education,
construction industries for prototyping purpose. The application area such as
automotive, aircraft, medical equipment’s requires very accurate and precise 3D printed
products without any defect.
iii. How did you come to know of this, did the people who have problem approach
you or you visualized it yourself?
We observed the common defects that occurs in 3D printed products by ourselves during
printing of products. Also, available literature related to additive manufacturing listed the
various types of defects occurred in 3D printing.
iv. Why is it important to solve it?
The accurate detection of faults in 3D printed components helps to eliminate the faults in
products. Development of Digital Twin (DT) setup for 3D printer fault detection help to
minimize the faults in 3D printing.

II. Baseline Data of the Identified location/Beneficiary:

3D printed products are widely used in manufacturing industries, education, medical,
construction, art and jewellery etc. Predominantly in manufacturing, preparing the component
prototype using the conventional way of mould formation is an expensive and time-consuming
process. So, such industries start using the rapid prototyping approach by using 3D printing
technology. The complex geometry parts with improvement in design in subsequent trials by
changing dimensions precisely can easily be manufactured using the rapid prototyping approach
of a 3D printer. These types of 3D printed products are subjected to a few defects (mentioned in
proof of concept and technology gaps sections). So, minimizing such defects is much beneficial
for
the rapid prototype manufacturing industries along with medical, education and construction
industries.
III. Proof of concept/prototype devised (if any):

In primary work, the faults are induced in the 3D printer by changing machine parameters and
machine structure. Table 1 shows the generated cubical box

Table 1. Fault in 3D printed product.

Sr.
3D product Fault Images
No.

1 Normal print of the object

2 Disturbing Bed Leveling

(Level up)

3 Disturbing Bed Leveling

(Level down)

4 Disturbing Extrusion
Temperature: (260 oC)

5 Disturbing Extrusion
Temperature: (180 oC)
 6 Disturbing Belt Tension

IV. Technology gaps & Suggested solution:

Many researchers come across the various faults that occur in the 3 D printed end product [1][2].
The common defects that occur in 3D printed products are warping, stringing, cracking, porosity
etc.[3].
At the initial stage, with the help of multi-sensor data, the different types of faults are identified
using the Data-Driven model [4]. The concept of Explainable AI is used to identify the reason for
the fault that occurred in 3D printed products. Digital twin [5] models can be implemented in
additive manufacturing to infer quality deviation during the manufacturing of products to
overcome this problem related to the quality of 3D printed products.
1. Review of Status:
Kadam et al. proposed layer-wise anomaly detection using pre-trained models and assembled
learning approach [1]. In this work, the author performed a comparative study of different pre-
trained models with classification algorithms for fault detection in 3D printers. The author stated
that applied concepts could also be used for real-time fault detections. Banadaki et al. proposed
the automatic quality grading system in additive manufacturing by using a deep learning
algorithm approach [6]. The model is able to classify the products based on quality grading of the
end product with an accuracy of up to 94%.

2. References:
[1] V. Kadam, S. Kumar, A. Bongale, S. Wazarkar, P. Kamat, and S. Patil, “Enhancing Surface
Fault Detection Using Machine Learning for 3D Printed Products,” Appl. Syst. Innov., vol. 4,
no. 2, 2021, doi: 10.3390/asi4020034.

[2] L. Yuan, “Solidification Defects in Additive Manufactured Materials,” Jom, vol. 71, no. 9, pp.
3221–3222, 2019, doi: 10.1007/s11837-019-03662-x.

[3] O. Abdulhameed, A. Al-Ahmari, W. Ameen, and S. H. Mian, “Additive manufacturing:

Challenges, trends, and applications,” Adv. Mech. Eng., vol. 11, no. 2, pp. 1–27, 2019, doi:
10.1177/1687814018822880.

[4] S. Sayyad, S. Kumar, A. Bongale, P. Kamat, S. Patil, and K. Kotecha, “Data-Driven Remaining
Useful Life Estimation for Milling Process: Sensors, Algorithms, Datasets, and Future
Directions,” IEEE Access, vol. 9, no. July, pp. 1–1, 2021, doi: 10.1109/access.2021.3101284.

[5] M. Singh, E. Fuenmayor, E. P. Hinchy, Y. Qiao, N. Murray, and D. Devine, “Digital twin:
Origin to future,” Appl. Syst. Innov., vol. 4, no. 2, pp. 1–19, 2021, doi: 10.3390/asi4020036.

[6] Y. Banadaki, N. Razaviarab, H. Fekrmandi, and S. Sharifi, “Toward Enabling a Reliable

Quality Monitoring System for Additive Manufacturing Process using Deep Convolutional
 Neural Networks,” pp. 1–8, 2020.

[7] G. L. Knapp et al., “Building blocks for a digital twin of additive manufacturing,” Acta
Mater., vol. 135, pp. 390–399, 2017, doi: 10.1016/j.actamat.2017.06.039.

3. Proposed Objectives:

i. To develop the Multi-sensor Data-Driven model for detecting a fault in 3D in Printer

ii. To build an Explainable AI model for creating a more explainable model for different
faults with accurate prediction.
iii.Infer the quality of 3D printed products using Digital Twin of 3D printer to minimize
the faults in 3D printed products.

4. Methodology (100 words):

The data is collected from multiple sensors such as vibration, sound, current and temperature
sensors. The collected data is extracted into the time, frequency, and time-frequency domain for
converting raw sensor data into meaningful data. The extracted features are mapped with the
respective faults. The more relevant features are selected from extracted features by using the
proper feature selection method. The selected features are provided to the machine or deep
learning model for classifying and detecting the type of fault. In the next stage, the Explainable AI
(XAI) is used to understand the cause of defects. Figure 1 shows the workflow for 3D printer
Digital Twin for fault diagnosis. XAI provides the explainability and interpretability to answer the
“black box” of machine learning to understand why a particular decision has been made. In the
final stage, cloud services are used to store real-time data in the cloud. In the cloud, by using
Digital Twin (DT), real-time fault prediction and its explanation are provided to minimize further
faults in 3D printer products.
 Figure 1: Workflow for 3D printer Digital Twin for fault diagnosis

The sensors will be mounted on various locations for testing purposes, and respective data will
be collected. Based on collected data, the proper location of sensors will be finalized. The data
will be collected by using the NI DAQ system by selecting the proper sampling frequency. The
collected data will be stored in the device. The normal and abnormal (faulty) data will be labelled
based on the type of fault. The classification algorithms will be used to check the training and
testing accuracy of the model after the complete data is considered altogether. For performance
evaluation, confusion matrices and ROC curves will be considered to select the best classification
algorithm available for various Machine Learning and Deep Learning algorithms. The pre-
processed labelled data is provided to the explainable AI model
 Figure 2 : shows the complete architecture of 3D printer Digital Twin (DT). The following
section explains the different subsystems used in the DT setup.
Digital Twin for Product (DT product):
For representing the product in the cyber domain, CAD/CAM tools are used. DT products
represent the parameters such as surface roughness, dimension, mechanical strength, etc. G/M
codes are used to carry geometry information like the process and geometry of the product. STL
files created by CAD software consist of information related to the geometry of the product, and
CAM software slices the product into multiple layers n it gives G/M codes related to the product.
G codes help in controlling the motion of machines like X, Y, Z axes control and filaments
extrusion rate etc. M code help to control machine parameters like temperature, stepper motor
acceleration etc.
Physical Twin for System (PT System):
DT product from design phase gives input to physical twin system of 3D printer (PT System) for
production purpose. This PT system considers factors such as the energy required, raw material
and inputs from DT products for the actual production of the product. Along with these PT
systems also consider parameters like flow rate, nozzle temperature, acceleration of stepper
motor and environmental parameters such as humidity, temperature, pressure etc. during
production
Multi-sensors:
Multi sensors collect the data from the DT system, PT system and Environment system to find the
interaction of parameters. Different sponsors like vibration, current, temperature humidity etc.,
are used to collect the data from these systems. The data is collected using Compact DAQ system.
Collected data helps to construct a stochastic function that carries tasks like; deviation between
PT product and DT product parameters, the deviation between DT product before manufacturing
PT product and continuous update of DT system.
Parsing of DT product:
Data collected from the sensor need to be parsed to its corresponding DT product parameters for
creating a reference library. G/M codes of DT products are used for parsing, which contains
information related to processing parameters such as acceleration of motor, temperature, etc,
along with product geometry information.
Feature Extraction
Sensor data is used to generate a reference library by using feature extraction from it. The parsed
sensor data extracts various time and frequency domain features such as energy, root mean
square values, peak to peak features, standard deviation, Kurtosis, skewness, mean- median
frequency, and signal to noise ratio.
Synchronization and Segmentation of features:
Extracted features synchronization and Segmentation is done based on parsed DT product
parameters. This helps to simplify the process of creating a reference library in the DT system
environment
Machine Learning algorithm:
To generate a reference library from DT products parsed parameters, we need to generate
clusters containing similar features using clustering/ outlier algorithms. We may use silhouette
coefficient for clustering purposes which check feature similarity with other clusters.
Fault Localization Algorithm:
The main purpose of the system digital twin is to detect and localize the fault of the physical
product. The reference library, which was created using a clustering algorithm, is used to detect
and localize DT products. The fault localization algorithm helps to find the deviation of DT
product parameters from its stored reference library parameters. Fault localization algorithm
Parses the DT products features run-time as well as after manufacturing physical product.
DT Update Algorithm:
The reference library of DT products needs to update regularly to keep the DT model alive.
Before updating the reference library, we need to check whether deviation or anomaly is
temporary due to system fault or due to machine parts degradation. The fault localization
algorithm needs to modify for keeping all deviating variables on track. Previous reference library
features need to be compared with the latest features drawn from the sensor data; if the
deviation is significant, we need to update the reference library to keep the DT model updated
and alive. Figure 3 shows the complete flow chart for creating a digital twin of system.
 Figure 3: flow chart for creating a digital twin of system

System Degradation Prediction Analysis:

For detecting system degradation, the digital twin model in to update regularly. DT products flow
rate may vary beyond the optimal range. This machine degradation may occur due to mechanical
degradation like stepper motor degradation over time, worn out or rollers, worn out of extruder
nozzle, etc. may cause a change in the extrusion flow rate.
Quality Inference Model with Explainable AI:
For generating a quality inference model, we need to build a quality deviation function which is
the function of sensor data, process parameters, product design parameters and environmental
parameters. Environment parameters affect physical system parameters, which leads to
deviation in a quality product. This deviation in quality due to environmental parameters also
affects sensor output data. Correlation between these parameters can be modelled using a
physical-based model derived from the first principle model. A data-driven approach can be used
with supervised learning algorithms for this complex multi-domain analysis. This quality
deviation function is also updated one DT model updates. The Explainable AI (XAI) is used to
understand the cause of defects. XAI provides the explain ability and interpretability to answer
the “black box” of machine learning to understand why a particular decision has been made. The
AI Model takes input from
the current task to make a recommendation or decision. Model structure can aim for direct
knowledge of model architecture (by using directly explainable models) or the black box models
to explain the model after studying model behaviour.
Sensor Positioning:
For building the Digital twin, we will be measuring the process parameters, Design parameters
and environmental parameters. Process parameters are generally the input parameters that will
generally be used for printing. Design parameters are generated using the cad file and software,
which is also the input parameter we consider for printing. Environmental parameters are the
variating taking place during the change in the input parameters, which are collected using
different sensors like Vibration sensors, Magnetic sensors, Acoustic sensors, current sensors,
temperature, and humidity sensors.
Vibration sensors: As fluctuating radial force continuously acts on the stator core of the stepper
motor during printing, we can say that stepper motors are a major source of vibration in 3D
printers. Thus, the vibration in each stepper motor is directly a result of various machine
parameters given as an input to the 3D printer. Accelerometers are devices that measure a
structure's vibration or acceleration of motion. So, to capture the vibration in a 3D printer, the
vibration sensors are mounted on stepper motors and the bed of the 3D printer.
Temperature: While printing, the filament is getting heated and converted into a molten stage.
The temperature in the room may have an even more drastic effect on certain products, including
ABS and other higher-temperature plastics. The difference in the temperature may cause the
material to cool quickly, which in turn produces defects in the printed part. The temperature
sensors are placed near the nozzle of the extruder and below the bed to monitor the temperature
and humidity of molten filaments
Current Sensor: Current sensor reveals the primary consumers of power in a 3D printer, such as
stepper motors, bead heaters, extruder heaters, fans etc. stepper motors power consumption
varies with its movement and speed. A 3D printer has a constant DC power supply. Hence, a
current sensor can be placed to monitor the varying current flow and capture the power
consumption in a 3D printer for each G/M code.

Preliminary work:

The induced faults are captured using multiple sensors such as vibration, sound and current
sensor.

Figure 4: Representation of normal and abnormal (low and high belt tension) condition
vibration, sound, and current sensor data.

Figure 4 shows the representation of normal and abnormal vibration, sound and current sensor
data for low and high belt tension conditions. It is clear that as belt tension is low, the intensity of
vibration and sound sensor increases, and the intensity of the current sensor decreases. For high
belt tension, the intensity of vibration, sound, and current sensor increases.
Figure 5: Anomalies detection of normalized vibration sensor data (a) Low belt tension (b) High
belt tension.

The data anomalies can be found based on normal and faulty data by defining the proper
threshold values, as shown in figure 5.

5. Work Plan (150 words - Please also provide activities schedule – Pert Diagram):

i. Phase wise work plan of action with time line and deliverables in tabular form
(Describe how the proposal includes a plan for pilot application or trial in a realistic
user environment of the technology/product, where the expected impacts to meet
end user needs may be demonstrated to the fullest feasible extent).

Time duration in months

Sl.
Activity
No.
2 4 6 8 10 12 14 16 18 20 22 24
Literature review and Project
1
formulation
Purchase of 3D printer Machine,
sensors and other required
2 software and hardware
Equipment’s
3 Preparation Digital twin setup

4 Trial Experimentation and Review

Analysis of collected data and
5 planning for the updating in the
system.
Experimentation (using different
6
material)
Analysis of Results (Clustering
7 algorithm and Fault localizing
algorithm)
 System Degradation Prediction
8
Analysis
9 Documentation of publications

ii. Technology Selection (State the criteria used for selection of technology for
addressing key problem(s) and the assessment of available technologies related to
the project)

Name of agency/institution/individual
The Source expert
Generated in-house by staff Symbiosis Institute of Technology,
Symbiosis International (Deemed
University), Pune 412115, India
Generated in-house by employing outside -
experts
Borrowed from an outside -
institution/expert
Modification of technology/know-how -
being used by the beneficiaries
Any other (please specify): -

technology mainly focused on the analytics part using techniques in the Artificial
Intelligence area.
iii. Technology Development/Adoption/Modification/Capacity Buiding – as
applicable (Provide information on the new R&D/adapted R & D to be carried out
for technology.
The R and D part will focuse on the
 Big Data analysis part,
 Machine Learning and Deep Learning algorithm part for model training and testing
 Cloud Services for real-time data storage and analysis.
iv. Institutions/places where detailed lab/field testing or experiments will be
carried out:
Symbiosis Institute of Technology, Symbiosis International (Deemed University), Pune
412115, India

v. Source of Technology:
Artificial Intelligence (AI) is a growing field in Industry 4.0. This technology helps to
diagnose and prognose the machine health condition. Due to the remarkable benefits of
AI techniques in manufacturing industries, Prognostics and Health Management (PHM)
researchers attracted to this technique.

vi. Mechanisms for Beneficiaries mobilization & Involvement:

(Please indicate how mobilization & participation of beneficiaries in the project work will
be ensured)
 Formation of new SHGs/technology user group or beneficiaries’ group for project
implementation and Involvement of existing SHGs through demonstration of the
usefulness of technology

8. Environmental, Legal and Ethical Issues: NA

9. Deliverables (the list below must correspond with and be derived from # 4, # 5 & # 6. Please
also indicate affordability of deliverables to the target beneficiaries):
Deliverable Mark √ Brief description
Product development/adaptation - -
The project provides the
development in the process to
improve the quality of the end
Process development/adaptation √
product by monitoring and
controlling the process of additive
manufacturing
Machine learning and deep learning
models are used to develop the
Technology package for development of the √ condition monitoring and Digital
Twin model.

project area
The research paper will be
Technology capability development,
training & √ published on the results obtained
from the adopted technology.
documentation (e.g. reports, papers,
articles,
technology manuals, patents)
The developed condition monitoring
for fault diagnosis can be
Scientific knowledge and/or data
generation implemented in future in other
critical application areas such as
automobile, medical, aerospace etc.
leading to technology development in
future
Other (Please specify) - -

10. Estimated Benefits:

Benefit Mark √ Brief description

The proper fault detection help
to decrease in product rejection
Economic (Cost-benefit √ rate.
analysis)
Employment generation - -
Promotes sustainable
Social √ development
Additive manufacturing is a
sustainable manufacturing
Environmental including √ approach. The identification
 and localization of faults in the
product reduce the rejection
rate.
potential for CDM benefits
Others (Please specify)

11. Self- sustainability of the project after SEED’s project support is over
(30 words):
Once the digital twin setup is developed for the 3D printer,
the project can easily sustain with minimum running cost.

12. Possibility of replication of the project in similar areas

(after the proposed technological solution is proven, how it
will be distributed? Involving state govt. for large scale
technology dissemination or via market or any other means –
any entrepreneur or business person involved in the work in
any manner?):
The replication of the technology is possible. Even this
concept can be implemented in various engineering and
health sectors. The technology can also be patented.
13. Suggest measurable indicators

S. No. Indicators (as applicable)

1 No. of faults identified during printing using classification

2 Fault classification accuracy in percentage

3 No. of faults eliminated due to implementation of Digital Twin

4 Cost impact due to elimination of faults

No. of organizations motivated and mobilized for replication of project

5 achievements

6 No. of publications produced (Title, Journal, issue, yr.)