The use of data and analytics in the manufacturing industry hasn’t been explored to its full potential despite advances in IoT, AI and big data. A huge part of the industry still relies on humans to maintain machines on the shop floors. The process is labour-intensive and leads to a loss in productivity and sales in the long run.
However, the industry is now taking baby steps towards the use of predictive analytics and machine learning to drive efficiency in machine maintenance. Arvind Mahishi, AVP and Vighnesh Subramanian, Senior Data Scientist at Tiger Analytics, talked about ML-powered digital twins at MLDS 2021.
Digital twin is an affordable, scalable and self-sustaining alternative in telling machine operators the exact condition of the equipment under their care. Mahishi said it takes in real-world data about a physical object or system as inputs and products as outputs to predict how the physical object or system will be affected by those inputs.
ML-Powered Digital Twin
Before ML started to revolutionise machine maintenance in a manufacturing setup, data was collected by reactive maintenance, scheduled maintenance, alarm history and root cause analysis; both time-consuming and insufficient. With AI and ML taking the centre stage, predictive and prescriptive maintenance have made a huge difference in the data collection quality.
The manufacturing industry is now relying on sensor data, temperature data, pressure data, airflow data, weather data etc to streamline the processes. Based on this data, they are productionising storage processes and building quality controls. The data enrichment process helps identify gaps, build parameters for data management, and create tools to access the quality of data ingested. Data is integrated with BI tools to create metadata for holistic data review on further data processing and analysis.
Talking about ML-powered digital twin roadmap, Mahishi said it begins with making an inventory of data from equipment and sensors, which helps with the overall infrastructure’s readiness assessment. Other steps include value complexity prioritisation, PoC and value demonstration, full-scale implementation and monitoring, followed by ongoing value measurement of the overall system. Based on the sensor data processing, a warning is created if the machine starts to show irregularities in the data collected, allowing time for the operator to carry out preventive repair.
Mahishi explained the use case of an Indian Steel manufacturing MNC to deploy a digital twin reliability model to improve predictive maintenance. He said they faced challenges such as data inadequacies, missing data, and broken pipeline. Moreover, the existing anomaly detection models were extremely time-consuming, required individual model maintenance and had no standardised framework.
Explaining the solution approach, Subramanian said they have collected sensors-based, near real-time information to execute a thorough analysis approach. The steps included:
- Initiation and data discovery: Extract raw data from equipment followed by data pre-processing and cleanup
- Data Processing: It involves regularising time stamps and resampling to the required time interval
- Data filtering: Removing data with low variance
- Model data set prep
- Anomaly detection: It involved testing multiple model frameworks, score built models on the entire dataset
- Forecasting: It identifies Impute gaps in time-series data resulting from machine learning being idle/tested. Tools such as ARIMA, random forest, ARIMAX, Smoothing were used
The final output reflecting system health is updated on a real-time dashboard. It further showcases results on anomalous behaviour, forecast of reliability and sensor-based root cause analysis. UI monitoring is crucial for the business owners to display the overall health status between the selected time window.
Highlighting the process’s business outcome, Mahishi shared that their solution has so far resulted in a saving of USD 2.75 million in terms of cost of maintenance saved and production loss averted. Currently, scalable models are being deployed in blast furnaces, crushers, fans, transformers, and similar equipment.