Unlike other industries such as retail, eCommerce or pharma, the role of AI and data science in the manufacturing industry is not widely known. The data generated in the manufacturing industry is hard to capture and therefore lags in leveraging AI in productivity and also moves the KPI. Ramesh Kumar, head of analytics at Tata Steel, spoke about how AI is being explored in the company and what are some of the challenges that come into picture while deploying AI. He is currently driving a large scale AI Implementation in manufacturing across Tata Steel.
Challenges Of Deploying AI In Manufacturing
Kumar shared four significant challenges that intrude the AI function in the manufacturing sector.
Data
Data is mostly generated from sensors deployed at various processes and functions in the manufacturing industry and plays a crucial role in deploying and generating data. It, therefore, requires careful installations, storage and auditing of data to source data in the manufacturing industry. To develop and deploy deep learning and machine models, it is the data collected from these sensors that play a crucial role. Further to generating data, it is essential to store, and the audit is carefully done to avoid any deviations in results.
People
People are the repository of knowledge, multiple key indications, processes and functions. They form a crucial asset to drive data science-based applications in the field. Kumar shared that most steel and manufacturing industries are located away from the nodal points and are often challenging to recruit or employ teams at these locations. Setting up a team is the first challenge.
To overcome this, Kumar shares that Tata Steel runs multiple training programs to acquaint people about various tools and technologies used in the process. For instance, to communicate data science terminologies with the ground-level operator to get the required deliverables from them can be a challenge. Therefore, these initiatives help in communicating data science in layman terms.
Process
The second challenge in the manufacturing industry is the process. It usually takes years to develop processes in the manufacturing sector. Data changes every few months. For instance, due to COVID, the manufacturing capacity at Tata Steel dropped drastically, which led to a decline in the data-generated through sensors. The primary data that was collected on which models were trained were of no use, and they had to be trained yet again.
External factors such as environment, raw materials, the property of coal and others might also affect the process and the final output. There are thousands of sensors that are installed across processes, the data from which are trained from historical data, but deviations in it may cause the models to deem irrelevant. This is where data preparation comes into the picture, which is extremely important. Further, any inherent change in the system, if the error is not corrected, will lead to unexpected results in the model.
Technology
Technology, again, is a challenge but not as severe as others. The availability of publically available libraries and tools make it easier for data scientists to explore and play around with data.
Use Cases In Tata Steel
Kumar shared that the data science function at Tata Steel runs on three principles — simplify, synergy and scale. Further explaining this, Kumar said that by ‘simplifying’ it means that data science should be communicated in a way for a common man to understand the technicalities. ‘Synergy’ is where data from different domains such as R&D, processes and machines come together to get into the entire working, whereas, the ‘scale’ is where one model should have the potential for horizontal deployment.
Talking about the use cases, Kumar shared that they use methods such as regression models in machine learning, predictive modelling, neural networks, gradient boosting, among others to generate insight from various data points. He further shared that they have models at every workstation, which are all stitched together to make it work and generate the desired results. He also shared that most of the models work on adaptive learning, i.e., they learn from the past error and keep updating taking care of any deviations in the future.
The connected machines and interdependency of the process in a complex manufacturing environment make it difficult for AI to find possibilities in the sector and measure its integrated impact. However, they are working with industry experts and even academics to bring about better applications of AI-enabled manufacturing.
