The International Society of Automation (ISA) reports that 5% of plant production is lost annually due to downtime. This translates to approximately $647 billion in global losses for manufacturers across various industry segments. The critical challenge is predicting maintenance needs to minimize downtime, reduce operational costs, and optimize maintenance schedules, according to NVIDIA Technical Blog.

LatentView Analytics

LatentView Analytics, a key player in the field, supports multiple Desktop as a Service (DaaS) clients. The DaaS industry, valued at $3 billion and growing at 12% annually, faces unique challenges in predictive maintenance. LatentView developed PULSE, an advanced predictive maintenance solution that leverages IoT-enabled assets and cutting-edge analytics to provide real-time insights, significantly reducing unplanned downtime and maintenance costs.

Remaining Useful Life Use Case

A leading computing device manufacturer sought to implement effective preventive maintenance to address part failures in millions of leased devices. LatentView’s predictive maintenance model aimed to forecast the remaining useful life (RUL) of each machine, thus reducing customer churn and enhancing profitability. The model aggregated data from key thermal, battery, fan, disk, and CPU sensors, applied to a forecasting model to predict machine failure and recommend timely repairs or replacements.

Challenges Faced

LatentView faced several challenges in their initial proof-of-concept, including computational bottlenecks and extended processing times due to the high volume of data. Other issues included handling large real-time datasets, sparse and noisy sensor data, complex multivariate relationships, and high infrastructure costs. These challenges necessitated a tool and library integration capable of scaling dynamically and optimizing total cost of ownership (TCO).

An Accelerated Predictive Maintenance Solution with RAPIDS

To overcome these challenges, LatentView integrated NVIDIA RAPIDS into their PULSE platform. RAPIDS offers accelerated data pipelines, operates on a familiar platform for data scientists, and efficiently handles sparse and noisy sensor data. This integration resulted in significant performance improvements, enabling faster data loading, preprocessing, and model training.

Creating Faster Data Pipelines

By leveraging GPU acceleration, workloads are parallelized, reducing the burden on CPU infrastructure and resulting in cost savings and improved performance.

Working in a Known Platform

RAPIDS utilizes syntactically similar packages to popular Python libraries like pandas and scikit-learn, allowing data scientists to speed up development without requiring new skills.

Navigating Dynamic Operational Conditions

GPU acceleration enables the model to adapt seamlessly to dynamic conditions and additional training data, ensuring robustness and responsiveness to evolving patterns.

Addressing Sparse and Noisy Sensor Data

RAPIDS significantly boosts data preprocessing speed, effectively handling missing values, noise, and irregularities in data collection, thus laying the foundation for accurate predictive models.

Faster Data Loading and Preprocessing, Model Training

RAPIDS’s features built on Apache Arrow provide over 10x speedup in data manipulation tasks, reducing model iteration time and allowing for multiple model evaluations in a short period.

CPU and RAPIDS Performance Comparison

LatentView conducted a proof-of-concept to benchmark the performance of their CPU-only model against RAPIDS on GPUs. The comparison highlighted significant speedups in data preparation, feature engineering, and group-by operations, achieving up to 639x improvements in specific tasks.

Conclusion

The successful integration of RAPIDS into the PULSE platform has led to compelling results in predictive maintenance for LatentView’s clients. The solution is now in a proof-of-concept stage and is expected to be fully deployed by Q4 2024. LatentView plans to continue leveraging RAPIDS for modeling projects across their manufacturing portfolio.

Image source: Shutterstock