6.4. C2000 Family

C2000 real-time MCUs are TI’s platform for motor control, power conversion and other real-time control systems. Running an AI model next to the control loop enables predictive maintenance and fault detection (arc fault, motor bearing fault, blower imbalance) on the same device.

Tiny ML ModelZoo 1.5.0 supports two C2000 CPU generations:

  • F28x devices with the 32-bit C28x DSP core (compiled with the TI C2000 code generation tools, cl2000).

  • F29x devices with the C29x DSP core (compiled with c29clang).

6.4.1. F28x Devices (C28x)

Device

NPU

Supported tasks

SDK example in C2000Ware

F28P55

Yes

Classification (generic, arc fault, motor fault, blower imbalance), regression, forecasting, anomaly detection

f28p55x

F28P65

No

Same as F28P55

f28p65x

F28003

No

Classification (generic, arc fault, motor fault, blower imbalance, ECG), regression, forecasting, anomaly detection

f28003x

F28004

No

Same as F28003

f28004x

F2837

No

Same as F28003

–

F28P551x

No

Same as F28003

–

F2838x

No

Same as F28003

–

F2837xS

No

Classification (generic, arc fault, motor fault, blower imbalance, ECG)

–

F2807x

No

Classification (generic, arc fault, motor fault, blower imbalance, ECG)

–

F28002x

No

Classification (motor fault, blower imbalance, ECG), regression, forecasting, anomaly detection

–

F280013

No

Classification (arc fault, motor fault, blower imbalance, ECG), regression, forecasting, anomaly detection

–

F280015

No

Same as F280013

–

F28E12

No

Regression, forecasting, anomaly detection

–

F28P65 is compiled with 64-bit floating-point support (--float_support=fpu64); all other F28x devices use fpu32.

6.4.1.1. New in 1.5.0 for F28x

  • F28E12 is a new device profile. Its models are the regression, forecasting and anomaly-detection models; there is no classification model for F28E12 in 1.5.0.

  • The Gen3 F28x devices F2838x, F28P551x and F28002x are now listed by the regression, forecasting and anomaly-detection models.

  • Device lists were tightened to the verified per-model support. For example, F280013, F280015 and F28002x are not targets of generic_timeseries_classification, and F2837xS and F2807x are classification-only.

  • Feature-extraction memory is estimated for F28P55, F28P65, F2837, F28003, F28004, F280013, F280015 and F28E12 (see Device Overview).

6.4.1.2. F28P55 (NPU)

F28P55 is the C2000 device with the TI NPU. Use an _NPU model and keep the default TINPU quantization so that the hardware NPU is used:

common:
  task_type: arc_fault
  target_device: F28P55
dataset:
  dataset_name: dc_arc_fault_example_dsk
training:
  model_name: CLS_1k_NPU
compilation:
  compile_preset_name: default_preset    # or compress_npu_layer_data, forced_soft_npu_preset

This is the configuration of examples/dc_arc_fault/config_dsk.yaml (with the preset spelled out). For the NPU rules see NPU Guidelines.

On F28P55 the default preset uses the hardware NPU for classification, regression and anomaly detection. Forecasting models are compiled for the CPU (ti-npu type=soft) in 1.5.0.

6.4.1.3. Other F28x Devices

All other F28x devices run the model on the C28x CPU. The same _NPU models can be used; they are compiled for the CPU:

common:
  task_type: generic_timeseries_regression
  target_device: F28E12
training:
  model_name: REGR_2k_NPU

6.4.2. F29x Devices (C29x)

Device

NPU

Supported tasks

SDK example in the F29H85x SDK

F29H85

No

Classification (generic, arc fault, motor fault, blower imbalance), regression, forecasting, anomaly detection

f29h85x

F29P58

No

Classification (generic), regression, forecasting, anomaly detection

–

F29P32

No

Classification (generic), regression, forecasting, anomaly detection

–

F29x devices run the model on the C29x CPU. Feature-extraction memory is estimated for F29H85. Several examples ship an F29H85 configuration, for example grid_fault_detection, mosfet_temp_prediction and generic_timeseries_anomalydetection.

common:
  task_type: generic_timeseries_classification
  target_device: F29H85

6.4.3. SDKs and Tools

Devices

SDK (version used by the deployment guides)

Where the AI examples are

F28x

C2000Ware 6.01.00.00

C2000Ware_6_01_00_00/libraries/ai/examples/<example>/<device>/

F29x

F29H85X-SDK 1.04

see Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check)

LaunchPads used in the deployment guides: LAUNCHXL-F28P55X, LAUNCHXL-F28P65X, LAUNCHXL-F280039C, LAUNCHXL-F280049C and the F29H85X-SOM-EVM.

You do not need the SDK to train and compile: ModelMaker uses the C2000 compilers found through TOOLS_PATH (see Installation). You need the SDK and Code Composer Studio to run the model on the board.

6.4.4. Typical Applications

Application

Example

DC arc fault detection

dc_arc_fault

Motor bearing fault and fan blower imbalance

motor_bearing_fault, blower_imbalance

Torque and temperature estimation

torque_measurement_regression, forecasting_pmsm_rotor_temp

Grid fault detection (F29H85)

grid_fault_detection

6.4.5. Next Steps