How Caterpillar Compresses Heavy Machinery Data Using Neural Networks

A method for shrinking massive amounts of sensor data from construction equipment into small, efficient packets for cheaper wireless transmission by using neural network training.

What it covers

This patent describes a system to reduce the cost of sending data from remote construction machines to a central office. It uses a neural network to group similar operational data points into 'neurons' within a multi-dimensional space. Instead of sending every single sensor reading, the machine only sends the coordinates of these neurons, the average distance of data points to those neurons, and how often each neuron was 'hit' by incoming data. This allows the remote office to reconstruct the machine's behavior without needing the raw, high-bandwidth data stream.

What it doesn't cover

• —Does not cover general-purpose data compression algorithms like ZIP or JPEG.
• —Does not cover supervised learning where the machine is trained on labeled data.
• —Does not cover transmission methods that do not rely on neuron-based model parameters.
• —Does not cover data processing that occurs entirely on an external server without local compression.

The clever bit

By using unsupervised learning to create a 'winning neuron' model, the system essentially creates a mathematical summary of machine behavior that adapts to the data, rather than using a fixed, rigid compression template.

Why it matters

In the mid-2000s, transmitting large amounts of telemetry data from remote job sites via satellite or cellular links was prohibitively expensive. This patent provided a way to maintain diagnostic visibility into heavy equipment health while drastically cutting the data volume, effectively making remote fleet management economically viable for companies like Caterpillar.

Real-world examples

• 1.Caterpillar Product Link remote monitoring systems
• 2.Remote diagnostic systems for heavy mining equipment
• 3.Fleet management software for construction machinery