Foreword
Anomaly detection is the process of identifying unexpected events/behavior in data, which differ from the norm. In a complex system comprising several rotating as well as non-rotating equipment, monitoring the health of such equipment via anomaly detection is important from the perspective of minimizing unplanned downtime, maximizing productivity and extending the life of the overall system. Additionally, being able to predict events/failures with a suitable lead time and take timely action serves to slow down the natural degradation of such equipment. This session aims to bring together ideas and methodologies on predictive maintenance from different perspectives, oil and gas to aircraft health monitoring and probably beyond. We look forward to some very interesting talks followed by a 15 min open mic session for the speakers and the audience to interact and exchange ideas.
About the Chair
Venkatesh Madyastha is currently a principal data scientist with the Data Science R&D team in Shell Bangalore, India, where his work includes using data to detect and predict anomalies in rotating as well as non-rotating equipment. Prior to joining shell, Venky worked at the Global Research Center, GE, India where he focused on data driven as well as model-based approaches to fault/event detection and prediction in such areas as GE power, GE water, GE aircraft engines, GE locomotives and healthcare. Venky holds a PhD in aerospace engineering from the Georgia Institute of Technology, Atlanta, Georgia, USA where his doctoral thesis work focused on Kalman filtering with feed forward neural networks for applications such as missile-target tracking, formation flying and obstacle avoidance. He has authored several international journal and conference publications, has several patents to his credit and has 13+ years of industrial experience in predictive modeling. Venky plays cricket, volleyball and enjoys running every now and then. He is a diehard fan of Pacino, Pesci and Denzel movies and enjoys good standup comedy where some of his favorites include, but are not limited to, Dave Chappelle, Seinfeld, Ricky Gervais, Frankie Boyle.
LinkedIn: https://www.linkedin.com/in/venkatesh-madyastha-9218035/
Contents
1. Welcome to the audience – Venkatesh Madyastha [3:15pm to 3:20pm]
2. Predictive AI to prevent unplanned shutdowns in Ammonia Plants – Arjun Bhattacharyya
a. Talk: 3:20pm to 3:40pm
b. Q&A: 3:40pm to 3:45pm
3. Airplane health management – Seema Chopra
a. Talk: 3:45pm to 4:05pm
b. Q&A: 4:05pm to 4:10pm
4. Novel method of mixture density estimation and its application on sensor limit estimation and process monitoring – Swanand Khare
a. Talk: 4:10pm to 4:30pm
b. Q&A: 4:30pm to 4:35pm
5. Anomaly detection of rotating and non-rotating equipment via mutual information – Asmi Rizvi Khaleeli
a. Talk: 4:35pm to 4:55pm
b. Q&A: 4:55pm to 5:00pm
6. Open mic discussion to exchange thoughts and ideas on predictive maintenance and concluding remarks – 5:00pm to 5:15pm
Speaker 1: Arjun Bhattacharyya [VP & GM SymphonyAI group, Bangalore, India]
About Arjun
Arjun Bhattacharyya is currently Vice-President & General Manager at SymphonyAI group, one of the fastest-growing private group of B2B AI companies to deliver next-generation AI-driven high value to retail, CPG, healthcare, industrial, media and financial services sectors. Arjun heads product development & operations including business in region for the group’s Industrial AI company. As part of his charter, he is building a large-scale Industrial AI platform with cloud & Edge Analytics to deliver asset and process performance solutions for major industry verticals that include mining, metals, oil & gas, chemicals and process manufacturing. Previous to this, Arjun spent several years in General Electric at progressively responsible leadership roles. His last role was that of VP Engineering at GE Power Digital where was responsible for building next generation of APM and Operations excellence suite of digital products for the Power industry. As the head of Analytics practice at GE Water, he delivered a highly successful digital remote monitoring & diagnostics platform InSight™ for the water industry. With over 20 years in the industry, Arjun’s competencies include Plant Systems & Process engineering, Advanced Controls, Optimization & Artificial Intelligence.
LinkedIn: http://www.linkedin.com/pub/arjun-bhattacharyya/3/99/69b
Talk: Predictive AI to prevent unplanned shutdowns in Ammonia Plants
Ammonia is amongst the most widely produced chemicals by tonnage. In this presentation, we will investigate an ammonia plant that lacked insights into predicting and avoiding process trips and was experiencing 2-3 unplanned production outages due to excess CO2 slip every year in the steam reforming process. Restarting the plant post-process trip takes 3-4 days and production losses in this period could be more than a million dollars. Predictive health management of underlying individual unit operations could not completely prevent these shutdowns demanding a process-wide solution. An AI-based process anomaly predictive solution was deployed that uses an adaptive digital twin of the overall process to predictively capture in real-time emergent bad actors & provides insights to operators to proactively take mitigating actions and prevent incipient outages. The AI model could detect the anomaly 6 to 12 hours in advance giving sufficient lead time to react in order to prevent a potential shutdown.
Speaker 2: Seema Chopra [Global Technical Leader – Data Analytics at Boeing Research and Technology, Bangalore, India]
About Seema
Dr Seema Chopra is working as Global Technical Leader – Data Analytics at Boeing Research and Technology, India. She has been recognized as part of Boeing Technical Fellowship and became Boeing India’s first ‘Associate Technical Fellow’ for Artificial Intelligence. Her current work includes developing next generation advanced health management technologies using analytics and big data platforms for commercial and defence applications. Prior to this role, she was with GE as a PHM (Prognostic Health Management) Technical Leader and was involved in the design and developing prognostic health management technologies to enable strategic growth for Condition Based Maintenance for Gas turbines. Seema earned her doctorate degree in Control engineering from IIT Roorkee, India and the focused area was to design Fuzzy Controller with Intelligent Design Approaches with reduced rule set.
Seema is certified Black belt – DFSS Lean Six Sigma and has 14 patents, 5 Trade secrets and 35+ publications in various International/National journals & conferences, 8 Technical reports and received several awards for leadership and technical expertise including PHM Expertise award from President & CEO, GE Power Gen Services and GE Impact award 2012 from CEO of GE, for volunteering on Mid-Day meal.
Seema has 17+ years of research experience in the area of advanced analytics solutions for different applications. Her research includes different areas like Continuous Analytics, Fault Diagnosis and Prognosis, Real time streaming, Big Data platforms, Data Mining, Machine learning, IVHM and Control system.
LinkedIn: https://www.linkedin.com/in/seema-chopra-18729338/
Talk: Airplane Health Management
Speaker 3: Swanand Khare [Asst. Prof. Department of Mathematics and the Centre of Excellence in Artificial Intelligence, IIT Kharagpur, Kharagpur, India]
About Swanand
Swanand Khare is an assistant professor at the Department of Mathematics and the Centre of Excellence in Artificial Intelligence at IIT Kharagpur. He obtained the M.Sc. and Ph.D. degrees from IIT Bombay in 2005 and 2011 respectively. He was a post-doctoral researcher at the University of Alberta, Canada from 2011 to 2014. His research interests include inverse eigenvalue problems, computational linear algebra, estimation and computational issues in applied statistics. He is a recipient of Excellent Young Teacher Award 2018 at IIT Kharagpur.
LinkedIn: https://www.linkedin.com/in/swanand-khare-056b7310/
Talk: Novel method of mixture density estimation and its application on sensor limit estimation and process monitoring
Expectation Maximization (EM) algorithm is one of the most popular algorithms used in the parametric density estimation in statistics. The main aim in the parametric density estimation is to estimate the parameters of the specified density from the evidence by maximizing the likelihood function. Thus, EM algorithm essentially seeks the absolute maximum of the likelihood function. In cases where the underlying density is a mixture density, the likelihood function assumes a complex form, thereby making the search of the absolute maximum a challenging task. It is often noted that the numerical algorithms, including EM algorithm, lead to a local maximum of the likelihood function rather than the absolute maximum. This leads to the parameter estimates which do not represent the data accurately and still are used in subsequent analyses which may lead to undesirable decisions. In order to overcome this issue, we propose a variant of EM algorithm which outperforms the existing EM algorithm in case of parameter estimation of mixture Gaussian densities. Our approach is based on the randomization of key steps, namely the expectation step and maximization step. We propagate a carefully sampled population of the points through each iteration of EM algorithm yielding a better estimate of the parameters when the algorithm converges. We show through synthetic data as well as industrial case study that our proposed approach outperforms the traditional EM algorithm implemented in standard packages such as Python.
We implement the proposed approach in one of the applications in predictive maintenance where it is very important to know the normal operation range of the various process variables. Oftentimes the process variables show different operation modes based on the overall operation modes at which the whole plant is operated. Thus, it is sensible to model the operation range of individual process variables with mixture Gaussian densities. In this application, we show that the proposed EM algorithm accurately estimates the parameters of the mixture Gaussian density and thereby enabling us to design the alarms of varying degrees of importance to indicate changes in the operation modes of individual process variables. The accurate estimation of parameters reduce the number of false positives making the alarm more realistic.
Speaker 4: Asmi Rizvi Khaleeli [Data Science Researcher, Shell Technology Center, Bangalore, India]
About Asmi
Asmi Rizvi Khaleeli is a Data Science Researcher at Shell India, where her work concentrates in providing data science solutions for predictive maintenance of company’s assets. She has 10+ years of experience in scientific computing, data analytics and machine learning in a variety of industries, including automotive, aerospace, and energy. Asmi holds an M.Sc. in Computational Sciences in Engineering from the Technical University of Braunschweig, Germany. She is a mother, a feminist and a music lover.
LinkedIn: https://www.linkedin.com/in/asmi-rizvi-09230aa1/
Talk: Anomaly detection of rotating and non-rotating equipment via mutual information
Predictive maintenance aims at reducing the number of failures occurred during production as any breakdown of machine, equipment or an abnormality in the process may lead to disruption for the product’s supply chain. Predictive maintenance is highly required and recommended for crucial components whose failure can cause severe function loss and hold high safety risk. Recent utilization of data and related techniques in predictive maintenance greatly improves system health condition and boosts the speed and accuracy in the maintenance decision making. In this work a normalized mutual information theory based solution is used for predictive maintenance of a system. The solution utilizes the normalized mutual information building non-linear correlations in pairs of system sensor data. The method takes a machine learning approach of comparing the correlations captured during stable period to those occurring in real time to detect anomalies and thence predict any upcoming failure.