Sitara Machine Learning SITARA-MACHINE-LEARNING (ACTIVE)

Description

Machine learning is a branch of Artificial Intelligence (AI) that uses algorithms to parse and learn from big data to make intelligent decisions. With great advancements in computing power and the availability of enormous labeled data sets, machine learning has the potential to impact the world to the same degree as the internet, or the transistor has done before.

There are two main parts of machine learning: training and inference. The training phase of machine learning is the algorithm’s learning phase. Once the algorithm has been trained to predict information reliably, its parameter values are frozen and deployed into the field to “infer” a result on any new data it receives.

The technology is at the center of Industry 4.0 or Industrial Internet of Things (IIOT) to make factories smart.  It is helping with inspecting parts on an industrial production line, enabling smart robots and machines to autonomously complete tasks with precision and speed, predicting the remaining useable life of a piece of equipment and in many such applications to improve productivity and reduce operation costs.

Running machine learning inference at the edge as opposed to in the cloud helps reduce latency, decrease network bandwidth requirements, and address the security and reliability concerns.

Sitara processors enable machine learning inference at the edge on Arm® Cortex-A cores using ArmNN, the open-source Linux software from Arm. Sitara AM57x devices, specifically, can also perform machine learning inference at the edge using Texas Instruments Deep Learning (TIDL) software framework, which runs on C66x DSP cores and EVE subsystems. Both TIDL software framework and ArmNN software come pre-installed on Processor SDK.

Advantages of using Sitara processors in smart factories:

• Integrated SoC with key industrial peripherals to provide the processing and connectivity needed for smart factory products:
  • Industrial protocol support.
  • Real-time control and processing with PRU and C66x DSP.
  • Accelerated deep learning for high performance and low power.
  • CSI-2 and parallel interface for camera connection.
  • Support for PCIe, USB3, SATA and more.
• Enhanced reliability with extended temperature & high voltage I/O, low Failure-In-Time (FIT) and high Power-On-Hours (POH) with device longevity.
• Lower system cost:
  • Enables machine learning inference at the edge for many machine vision applications.
  • Enables predictive maintenance at the edge in field, controller or operator-level devices.
• Lower system power, < 5W.
• Low latency.
• The same Processor SDK can be used with all Sitara devices.

Example applications where Sitara machine learning can be used:

• Industrial robots for automated sorting.
• Vision computer & optical inspection.
• Automated guided vehicle.
• Smart HMI.
• Optimize equipment settings.
• Track, identify and count people and objects
• Extracting useful information gathered from the data aggregator.
• Predictive Maintenance (PdM) (identifies anomalies, machine wear and expected lifetime of equipment).
• Classify and recognize specific sounds/audio patterns.

How to evaluate Sitara machine learning:

Below are the hardware options and software needed to begin evaluating Sitara machine learning solutions.

Step 1
Choose hardware:

• BeagleBone AI- Low cost machine learning evaluation module that is supported by the community.
• AM574x IDK- Highest performance evaluation with dedicated deep learning accelerators and 2x C66x DSPs.
• AM572 EVM- Evaluation of machine learning with 2x C66x DSPs.
• AM571 IDK - Evaluation of machine learning with 1x C66x DSP.
  

Step 2

Download Processor SDK Linux for AM57x:

• http://software-dl.ti.com/processor-sdk-linux/esd/AM57X/latest/index_FDS.html

Step 3

Review the following machine learning reference design demonstrating TIDL software framework on a Sitara AM57x processor:

Machine Learning Inference for Embedded Applications reference design

Software

TI Deep Learning (TIDL) software framework:

TIDL software framework leverages a highly optimized neural network implementation on TI’s Sitara AM57x processors, making use of hardware acceleration on the device. TIDL is a set of open-source Linux software packages and tools that enables offloading of deep learning inference to the Embedded Vision Engine (EVE) subsystem, the C66x DSP subsystem, or both. TIDL software is available as part of TI’s free Processor SDK: 

• Go to the Processor SDK download page (includes TIDL)
• View the TIDL Getting Started Guide for Processor SDK
• Review the TIDL demonstration documentation
  
  

Other Helpful Resources:

Bringing machine learning to embedded systems (white paper)