Problem
Domin Fluid Power, a leading manufacturer of electro-hydraulic actuators, faced a challenge in monitoring the health of their pumps deployed at customer sites. Traditionally, they relied on centrally trained machine learning models to analyse sensor data and detect abnormal pump behaviour. However, this approach had limitations:
- Data Privacy Concerns: Customer data resides on their infrastructure, raising privacy concerns about transferring it to a central location for training.
- Data Transfer Costs: Uploading large volumes of sensor data from geographically dispersed locations can be expensive.
The absence of real-time, on-site analysis constrained Domin Fluid Power’s ability to predict and prevent pump failures, which could lead to costly downtime.
Solution
Domin partnered with OctaiPipe to address these challenges and implemented a federated machine learning approach emphasising privacy preservation. OctaiPipe’s innovative FL-Ops approach allowed Domin to create, deploy, and manage lightweight Convolutional Neural Network (ConvNet) machine learning models directly on each customer’s pump at the network edge.
The federated edge AI infrastructure was designed specifically to ensure data privacy. Training data never left the customer’s premises, alleviating their concerns about data misuse and potential privacy breaches. This approach not only safeguarded sensitive customer data but also significantly cut down on data transfer costs.
OctaiPipe’s Impact
- Improved anomaly detection: 100x compared to the naive/dummy model.
- Enhanced model performance: Achieved a 2% accuracy gain in anomaly detection compared to the centrally trained model, specifically reflecting recall, which is an attribute of accuracy.
- Reduced data transfer cost: Achieved a 5x reduction after down-sampling and compression.
Domin’s adoption of OctaiPipe’s federated learning approach led to significant advancements in pump monitoring. The federated ConvNet model provided a remarkable enhancement in anomaly detection, outperforming the baseline model. The method not only improved the overall performance of the model but also drastically reduced data transfer costs by processing data locally.
About OctaiPipe
OctaiPipe is a federated edge AI infrastructure, delivering tailored solutions for Edge AIoT devices that challenge the status quo of traditional centralised cloud infrastructures. By setting new standards in AI deployment at the edge, OctaiPipe is not just enhancing the performance of Edge AI systems but also strengthening security and reducing costs. Our innovative approach enables data scientists to develop distributed solutions that are both scalable and efficient.
OctaiPipe’s Impact on Edge AI for IoT:
- Enhances IoT data performance via continuous learning, increasing by up to 2x the accuracy compared to single-device machine learning models, ensuring that insights and decision-making are based on the most reliable information.
- Offers cost-effective Edge AI for IoT by reducing costs, notably slashing cloud CPU costs by >100x, making advanced AI accessible to a broader range of IoT applications.
- Facilitates learning across 10,000+ devices without data centralisation, promoting a scalable approach that respects privacy and efficiency.
- Boosts AI resilience with less cloud and network dependence, significantly fortifying systems against AI and data attacks through secure GRPC encryption.
- Accelerates AI deployment for 500 devices in 3 hours, demonstrating a 10x reduction in wall-to-wall training times through parallelisation.
- Reduces learning times by 90%, streamlining the model development process and swiftly adapting to new data and contexts.
