fan_blade_fault_classification

Fan Blade Fault Classification

---

From collecting a dataset to deploying the model on a C2000 device

-Adithya Thonse, Shamik Basak, Tushar Sharma, Fasna Sharaf

---

Example showcasing 4 different Fan Blade faults (Blade Damage, Blade Imbalance, Blade Obstruction, Normal)

Involves using TI toolchains to collect the data and train the model and preview it live.

---

⚠️ Device Support: While this documentation focuses on the F28P55x, the following devices are also fully supported:

• CC1312
• CC1314
• CC1352
• CC1354
• CC2755
• CC35X1

Check the config_<device>.yaml files for device-specific configurations.

---

This application example showcases the following story:

• Loading your F28P55x device with a Live Data Capture Code
• Collecting a dataset on Model Composer
• Training an AI Model and Compiling it:
  • on Edge AI Studio - Model Composer
  • on Tiny ML Modelmaker
• Deploying the model on F28P55x device and visualising it in Live Preview Mode
• Other Tiny ML Modelmaker Features
• Release History

Capturing Data

• This example showcases connecting the F28P55x launchpad with ADXL355-PMDZ Accelerometer
  • 7 connections need to happen from the sensor to the launchpad
  • Accelerometer Pinout
  • 
  • F28P55 LaunchPad Pinout
  • 
  • Connections:

Accelerometer Sensor	F28P55 Launchpad
Pin 1 (Chip Select)	IO.27(J6:59)
Pin 2 (MOSI)	IO.61(J6:54)
Pin 3 (MISO)	IO.60(J6:55)
Pin 4 (SCLK)	IO.24(J5:47)
Pin 10 (Data Ready)	IO.16(J6:51)
Pin 11 (D GND)	Any Ground Pin
Pin 12 (D VDD)	Any 3p3 V Pin

• Extract and load eAI_data_acq_dap_f28p55x.zip from CCS Studio onto the F28P55x Launchpad.

  • This enables the device to communicate with the Edge AI Studio - Model Composer GUI
  • Go to the TI's Edge AI Studio Model Composer and start your new project (Desktop/Web version are both okay)
  • On the top: Options --> Serial Port Settings --> Click on the appropriate COM port where the launchpad is connected
    • Once selected, a small dialog at the bottom should say "Hardware Connected"
  • In the Capture section
    • Capture & Display Live Data --> Drop down Sensor: --> Vibration
    • Choose the other options as per preference. The default sampling frequency is 4kHz
    • Click on Start Capture
    • The data captured should look like this:
    • 

• In this manner, you can collect data for all your cases.

• This example has 4 classes: Normal, Blade Damage, Blade Imbalance, Blade Obstruction

  

• The entire demo setup looks somewhat like this:

• 

Dataset

• The dataset collected with this setup at TI can be found here

Training the Model

• You can use the Model Composer GUI or Tiny ML Modelmaker to train your model and compile it.
• In Model Composer, you can play with different settings that can get your model to accuracy suited to you
• For Tiny ML Modelmaker, you can use this config file and run it as follows from the tinyml-modelzoo directory
  • run_tinyml_modelzoo.sh fan_blade_fault_classification/config.yaml
• This configuration file should give you a 100% accuracy

Live Preview

• This should give you a compiled model in the tinyml-modelmaker/data/projects/fan_blade_fault_dsi/run/<time_stamp>/<model_name>/compilation/artifacts

• And few feature extraction related artifacts are a part of: tinyml-modelmaker/data/projects/fan_blade_fault_dsi/run/<time_stamp>/<model_name>/training/quantization/golden_vectors

• This model is ready to be deployed into the CCS project with the application code of your choice.

• You can refer to this for how to plug and play these artifacts.

• To run in Live Preview Mode, as shown in the TIREX above, a sample project with just 1 sensor is attached here: eAI_mfd_eval_f28p55x.zip

• You should be able to see in the Live Preview Mode in this manner:

• 

Available Default Configurations For Each Device Family

• config.yaml - F28P55
• config_CC1312.yaml - CC1312
• config_CC1314.yaml - CC1314
• config_CC1352.yaml - CC1352
• config_CC1354.yaml - CC1354
• config_CC2755.yaml - CC2755
• config_CC35X1.yaml - CC35X1

Other Tiny ML Modelmaker Features

• PCA graphs for feature extraction on Train and Val data: Clear cluster separation shows that the dataset classes are classifiable pretty well with the feature extraction method chosen in the config

• 

• 

• Post Training AUC-ROC Curve indicates a perfect score (1.00) --> indicates model is doing excellently well at classification!

  • 

• Histogram of class scores between each class combination shows good separability --> meaning that the AI model trained is able to distinguish between classes very clearly

  • 

Release History:

[26-Mar-2025]: Compatible with v1.0 of Tiny ML Modelmaker