electrical_fault

Electrical Fault Detection and Classification: Focus on solving multicollinearity problem

- Tushar Sharma, Adithya Thonse, Fasna S

---

Overview

The Electrical Fault Dataset is a multivariate time series dataset. It is obtained by modelling a 3 phase - transmission line of power system on MATLAB Simulink. Three-phase systems or electrical power systems often use three sinusoidal voltages (Va, Vb, Vc) that are phase-shifted by 120 degrees. Each phase also carries a current (Ia, Ib, Ic). The dataset is designed to find faults in transmission line and classify the type of fault using the line voltage and current.

The fault can be Line-to-line, Line-to-ground, Line-to-line-to-ground and more. Line-to-line (LL) fault is a fault between two phase conductors (e.g., A-B). This typically appears as a short-duration high-energy event affecting particular frequency bins. Line-to-ground (LG) fault: a fault between a phase conductor and ground. Pattern differs from LL faults and can be identified using combinations of voltage and current measurements. LLG / LLLG: multi-conductor faults involving two or three lines and possibly ground. These create distinct signatures across voltage and current channels.

Users can choose between two dataset options, each processed by a different script to produce a distinct output file:

• 2-class dataset (detect_dataset.xlsx): Processed by electrical_fault.py to create electrical_fault_dataset.zip for fault detection (binary classification - fault vs no fault)
• 6-class dataset (classData.csv): Processed by electrical_fault_6class.py to create electrical_fault_6class_dataset.zip for fault type classification (6 fault types based on G,C,B,A combinations)

There are 6 measurable parameters/variables - (Va, Vb, Vc, Ia, Ib, Ic) i.e the voltage and current of three phases. There are two dataset files present in the compressed zip: electrical_fault_raw.zip

• detect_dataset.xlsx : which finds if there is a fault or not. Target = [Output (S)]
  • There is one target only i.e. Output (S) which has two unique values (0, 1) denoting fault and no fault.
• classData.csv : which classifies the type of fault Target = [G, C, B, A]
  • There are 4 target variables i.e G (Ground), C (Node C), B (Node B), A (NodeA). The value of each target is either 0 or 1.
  • Examples [G, C, B, A]:
    • [0, 0, 0, 0] means No Fault
    • [0, 1, 1, 0] means LL Fault btw Node B and Node C
    • [0, 1, 1, 1] means LLL Fault btw all Nodes
    • [1, 0, 0, 1] means LG Fault btw Ground and Node A
    • [1, 0, 1, 1] means LLG Fault btw Node A, Node B and Ground
    • [1, 1, 1, 1] means LLLG Fault btw all Nodes and Ground

Depending on the dataset chosen, users can either detect whether there is an electrical fault (binary classification) or classify the type of fault (6-class classification).

Downloading dataset

Users can prepare the zipped dataset using either of two python scripts, depending on which dataset they want to use:

For 2-class dataset (fault detection):

cd examples/electrical_fault
python electrical_fault.py

This creates electrical_fault_dataset.zip

For 6-class dataset (fault type classification):

cd examples/electrical_fault
python electrical_fault_6class.py

This creates electrical_fault_6class_dataset.zip

The path to the appropriate zipped dataset file should be mentioned in configuration yaml under dataset.input_data_path, make sure it matches the script you ran.

dataset:
    input_data_path: 'examples/electrical_fault/electrical_fault_dataset.zip'   # OR 'examples/electrical_fault/electrical_fault_6class_dataset.zip'

Usage in Tiny ML ModelZoo

This zipped dataset is designed to work with Tiny ML ModelZoo. Run the modelzoo with the yaml configuration using the below code.

run_tinyml_modelzoo.sh examples/electrical_fault/config.yaml

1. run_tinyml_modelzoo.sh is the script to run modelzoo. It take two required arguments.
2. examples/electrical_fault/config.yaml path of the yaml configuration to run

The users can configure the yaml configuration to change parameters related to data preprocessing feature extraction, training, testing, model and model compilation. In this example, we will configure the parameters of feature extraction.

What if Multicollinearity in your dataset ?

Multicollinearity is a condition when independent variable posses a linear relationship with one or more than one independent variable. Raw voltage and current channels can be strongly correlated (for example, Vc may be nearly a linear combination of Va and Vb in some operating conditions). In the time domain this redundancy can confuse simple models. This gives rise to poor prediction of weights of ML model during training.

Transforming to the frequency domain (FFT) captures how energy is distributed across frequencies, and binning groups similar spectral regions together. Spectral features often decorrelate spatially correlated channels because they focus on frequency content rather than instantaneous amplitude. After binning and optional log-scaling, the model sees compact, lower-dimensional features that are more robust to correlated input channels.

In this electrical fault dataset, the independent variables are current and voltages of Phase A, B, C. If we check the collinearity between these variables. We find that Multicollinearity exists between some independent variables.

Independent Variable	Ia	Ib	Ic	Va	Vb	Vc
Ia	1.00	-0.49	-0.45	0.23	0.69	-0.94
Ib	-0.49	1.00	-0.54	-0.95	0.26	0.72
Ic	-0.45	-0.54	1.00	0.74	-0.94	0.17
Va	0.23	-0.95	0.74	1.00	-0.52	-0.51
Vb	0.69	0.26	-0.94	-0.52	1.00	-0.46
Vc	-0.94	0.72	0.17	-0.51	-0.46	1.00

To solve the problem of Multicollinearity, we can do one or more of the following:

• Remove the highly correlated features
• Perform feature extraction to do dimensionality reduction of features

Feature Extraction is the solution

In this example we will explore the method to do dimensionality reduction using FFT and BINNING of features. The data preprocessing feature extraction section of yaml configuration can be used to configure it.

When modifying the configuration file to disable feature extraction (as illustrated below) in modelmaker, you'll encounter an error that prevents proper model training with effective hyperparameters. This issue stems from multicollinearity in the data.

data_processing_feature_extraction:
   feature_extraction_name: Custom_Default
   feat_ext_transform: []

Now, with this information lets see how feature extraction will effect this error.

• FFT related options: FFT_FE, FFT_POS_HALF, DC_REMOVE, ABS
  1. FFT_FE is used to perform fft on a frame
  2. FFT_POS_HALF takes the 1st half of the fft which is symmetrical from middle
  3. DC_REMOVE removes the DC component of the FFT
  4. ABS takes the magnitude of the real and imaginary values of FFT
• Binning related options: BINNING
  1. BINNING performs the binning of magnitude of fft values
• Other options: LOG_DB, CONCAT
  1. LOG_DB takes the log of binned values
  2. CONCAT does concatenation of current features with features from previous frames

Example configuration walkthrough (numbers)

• frame_size = 256: using 256-sample frames. If the sampling rate is 2048 Hz, this corresponds to a time window of 125 ms.
• After FFT, positive half of spectrum length is frame_size/2 = 128 bins.
• feature_size_per_frame = 32 with binning: each feature aggregates 128 / 32 = 4 FFT bins (simple uniform grouping).
• num_frame_concat = 4: the final model input has 32 * 4 = 128 values per channel per example.

We have to add FFT and few more transforms in this transform variable.

data_processing_feature_extraction:
    feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'ABS', 'LOG_DB', 'CONCAT']

Configuration 1: FFT

Next we will define our features shape.

data_processing_feature_extraction:
    feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'ABS', 'LOG_DB', 'CONCAT']
    
    frame_size: 256
    feature_size_per_frame: 128
    num_frame_concat: 1

1. frame_size: slices frame from the dataset of size frame_size
2. feature_size_per_frame: size of binned features from one frame
3. num_frame_concat: number of frames used for concatenating features

After doing the above changes in yaml configuration file. Run the modelzoo again for this dataset.

run_tinyml_modelzoo.sh examples/electrical_fault/config.yaml

You can see that, you don't encounter any error during modelmaker run. This is because the feature extraction was succesfully able to mitigate the mulitcollinearity problem. This will resolve the error of to train the model properly with good hyper parameters.

Configuration 2: FFT + BIN

Another feature extraction is to perform FFT with Binning. For this, we need to add BINNING to transforms. The feature size for each frame would become half of the frame size and can be reduced further based on feature_size_per_frame selected. So, yaml configuration would look like.

data_processing_feature_extraction:
    feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'BINNING', 'ABS', 'LOG_DB', 'CONCAT']
    
    frame_size: 256
    feature_size_per_frame: 32
    num_frame_concat: 4

The first configuration employed individual frames (num_frame_concat: 1), while the second configuration utilized a sequence of 4 frames (num_frame_concat: 4). To decrease the model's size, you can lower the number of frames processed. For instance, setting num_frame_concat to 2 in the second configuration would approximately halve both the model size and processing time on your device.

Performance on device

We benchmarked the performance of the CLS_ResCat_3k model. The device used is F28P55x which comes with a HW accelearator (TINPU) to give low latency performance on ML models. Numbers are provided for running the model on NPU & CPU. Here both the configuration of Feature extraction produces the same architecture of model, so the model performance will be same. We clubbed the two configuration as 'with Feature Extraction'.

Configuration	AI Model Cycles	Inference Time (us)	Flash Usage (B)	SRAM Usage (B)
NPU (without Feature Extraction)	Bad Training	Bad Training	Bad Training	Bad Training
NPU (with Feature Extraction)	125509	836.73	3175	3846
CPU (with Feature Extraction)	500860	3339.07	2995	4992

---

Update history: [29th Dec 2025]: Compatible with v1.2 of Tiny ML Modelmaker [12th Mar 2025]: Compatible with v1.0 of Tiny ML Modelmaker