ECG Anomaly Detection at the Edge using MAX78000
-
install pyenv
https://github.com/pyenv/pyenv#set-up-your-shell-environment-for-pyenv -
Creating the Virtual Environment
$ cd <your/project> $ git clone --recursive https://github.com/analogdevicesinc/ai8x-training.git $ cd ai8x-training $ pyenv local 3.11.8 $ python -m venv .venv --prompt ai8x-training $ echo "*" > .venv/.gitignore $ source .venv/bin/activate (ai8x-training) $ pip3 install -U pip wheel setuptools (ai8x-training) $ pip3 install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu121 (ai8x-training) $ pip3 install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/rocm5.7 (ai8x-training) $ deactivate $ cd <your/project> $ git clone --recursive https://github.com/analogdevicesinc/ai8x-synthesis.git $ cd ai8x-synthesis $ pyenv local 3.11.8 $ python -m venv .venv --prompt ai8x-synthesis $ echo "*" > .venv/.gitignore $ source .venv/bin/activate (ai8x-synthesis) $ pip3 install -U pip setuptools (ai8x-synthesis) $ pip3 install -r requirements.txt -
Embedded Software Development Kit (MSDK)
Prerequisites
$ sudo apt update && sudo apt install libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-shm0 libxcb-util1 libxcb-keysyms1 libxcb-randr0 libxcb-render-util0 libxcb-render0 libxcb-shape0 libxcb-sync1 libxcb-xfixes0 libxcb-xinerama0 libxcb xkb1 libxcb1 libxkbcommon-x11-0 libxkbcommon0 libgl1 libusb-0.1-4 libhidapi-libusb0 libhidapi-hidraw0Download MSDK installer
Final Check
After a successful manual or MSDK installation, the following commands will run from on the terminal and display their version numbers:
arm-none-eabi-gcc -v arm-none-eabi-gdb -v make -v openocd -vgen-demos-max78000.sh and gen-demos-max78002.sh will create code that is compatible with the MSDK and copy it into the MSDK’s Example directories.
In this project, I will explore the fascinating field of electrocardiogram (ECG) signal processing and classification, leveraging the power of a Convolutional Neural Network (CNN) Autoencoder. My selected dataset for this endeavor is the PTB Diagnostic ECG Database, a carefully curated collection of ECG signals designed specifically for diagnostic purposes. My main goal is to build a tiny Autoencoder model tailored to the task of detecting abnormal electrocardiogram (ECG) signals at the Edge using MAX78000.
The following sections will present the abnormal detection system for the edge devices. The designed model is trained, deployed, and verified on MAX78000 EVkit.
The PTB Diagnostic ECG Database is a collection of 14,552 ECG recordings sourced from Physionet's PTB Diagnostic Database. These ECG signals are categorized into two classes: normal heartbeats and those affected by cardiac abnormalities. The dataset is sampled at 125Hz, providing high-resolution data for in-depth analysis.
Let's delve into some essential details about the PTB Diagnostic ECG Database:
- Number of Samples: 14,552
- Number of Categories: 2
- Sampling Frequency: 125Hz
Remark: All the samples are cropped, downsampled and padded with zeroes if necessary to the fixed dimension of 188.
Data Source: Physionet's PTB Diagnostic Database
Data Files: This dataset consists of a series of CSV files. Each of these CSV files contain a matrix, with each row representing an example in that portion of the dataset. The final element of each row denotes the class to which that example belongs.
The data loader handles all the data related tasks for proper model training such as raw data pre-processing, windowing, type casting etc. The data loader implementation is in the training repository, (datasets/sampleecg.py). Its functionality is described in the following sections.
The data loader also handles the train-test data splitting operations. In an unsupervised learning approach, the training set only includes normal samples without any arrhythmia. 20% of the normal data is also kept for the validation set and contributes to the testset as well. The testset includes all kinds of arrhythmia heartbeat.
For autoencoder model training, the data samples will have a label equivalent to the input signal since an MSE loss will be used and the aim of the model is to reconstruct the input as closely as possible. For the evaluation mode where anomaly detection performance is of interest, each data sample will have a binary label representing whether the raw data session is a normal case or includes any arrhythmia.
The neural network architecture is called an autoencoder. Autoencoders learn to reconstruct their input at their output. The autoencoder is trained so that the maximum information is kept after dimension reduction through encoding and has the minimum reconstruction loss (RL) after decoding. Therefore, the encoder and decoder structures are forced to learn the best encoding-decoding scheme in the training process. Autoencoders are well suited to an unsupervised learning approach using the normal signal reconstruction and provide a high RL when a arrhythmia case is encountered that deviates from normal data.
The model implementation file is in the training repository, (models/ai85net-autoencoder-ecg.py).
The input to the autoencoder is a 2D tensor of shape (187, 1). There are 187 channels each of width 1. CNN filters work depthwise across channels, so at the input layer, each filter is looking at a single axis at a time. The output data format is a 1D tensor of length 187 (1 axes × 187 samples). The reason is that the output layer is a fully connected layer, not a convolutional layer.
The use of a fully connected layer for the output layer (and several internal hidden layers) is to accommodate the fact that 1D transpose convolutions are not yet available in the target software stack. Transpose convolutions are generally used in the decoder to reconstruct the compressed representation of the data that the decoder received. Convolutional layers are used in the encoder to reduce the number of total parameters in the model.
The model training script is also available in the training repository, (scripts/train_autoencoder_ecg.sh).
==> Best [Top 1 (MSE): 0.01245 Params: 55456 on epoch: 368]
Saving checkpoint to: logs/2024.11.04-144418/qat_checkpoint.pth.tar
--- test ---------------------
810 samples (32 per mini-batch)
Test: [ 1/ 26] Loss 0.012738 MSE 0.012738
Test: [ 2/ 26] Loss 0.013812 MSE 0.013812
Test: [ 3/ 26] Loss 0.013366 MSE 0.013366
Test: [ 4/ 26] Loss 0.014103 MSE 0.014103
Test: [ 5/ 26] Loss 0.013647 MSE 0.013647
Test: [ 6/ 26] Loss 0.013966 MSE 0.013966
Test: [ 7/ 26] Loss 0.013821 MSE 0.013821
Test: [ 8/ 26] Loss 0.013421 MSE 0.013421
Test: [ 9/ 26] Loss 0.013230 MSE 0.013230
Test: [ 10/ 26] Loss 0.013219 MSE 0.013219
Test: [ 11/ 26] Loss 0.013406 MSE 0.013406
Test: [ 12/ 26] Loss 0.013289 MSE 0.013289
Test: [ 13/ 26] Loss 0.013554 MSE 0.013554
Test: [ 14/ 26] Loss 0.013605 MSE 0.013605
Test: [ 15/ 26] Loss 0.013698 MSE 0.013698
Test: [ 16/ 26] Loss 0.013600 MSE 0.013600
Test: [ 17/ 26] Loss 0.013713 MSE 0.013713
Test: [ 18/ 26] Loss 0.013680 MSE 0.013680
Test: [ 19/ 26] Loss 0.013739 MSE 0.013739
Test: [ 20/ 26] Loss 0.013638 MSE 0.013638
Test: [ 21/ 26] Loss 0.013691 MSE 0.013691
Test: [ 22/ 26] Loss 0.013681 MSE 0.013681
Test: [ 23/ 26] Loss 0.013818 MSE 0.013818
Test: [ 24/ 26] Loss 0.013912 MSE 0.013912
Test: [ 25/ 26] Loss 0.013826 MSE 0.013826
Test: [ 26/ 26] Loss 0.013875 MSE 0.013841
==> MSE: 0.01384 Loss: 0.014
Log file for this run: ai8x-training/logs/2024.11.04-144418/2024.11.04-144418.log
copy ai8x-training/logs/2024.11.04-144418/qat_checkpoint.pth.tar to ai8x-synthesis/trained/ai85-autoencoder-ecg-qat.pth.tar
The model quantize script is also available in the synthesis repository scripts/quantize_autoencoder_ecg_max78000.sh
The trained autoencoder model will generate an output with the input signal shape, and is trained to reconstruct the normal signal as closely as possible. A small post-processing step is required to deploy the model in the arrhythmia detection system: The system should mark some inputs as anomalies using the model output.
The basic principle of this post-processing step is to use the reconstruction loss (RL) level to detect any arrhythmia. Therefore, a pre-determined, learned threshold is needed for the decision boundary. Using training samples' RL percentiles is practical and a sample evaluation script is also available in training repository, (notebooks/ECG_AutoEncoder_Evaluation.ipynb) and demonstrates these post-processing and performance evaluation steps.
Several performance metrics such as balanced accuracy (average of True Positive Rate and True Negative Rate) and False Positive Rate, F1 score, etc. are evaluated in this notebook. Some evaluation plots for the Physionet's PTB Diagnostic Database are:
This file can be generated using the scripts/evaluate_autoencoder_ecg.sh script in the training repository.
(ai8x-training) ai8x-training$ scripts/evaluate_autoencoder_ecg.sh
Ignore this error
AttributeError: 'Namespace' object has no attribute 'quantize_eval'
Copy file sample_sampleecg_forevalwithsignal.npy in the project folder to tests/sample_sampleecg_forevalwithsignal.npy in the synthesis repository.
The file located at synthesis repository, (networks/ai85-autoencoder-ecg.yaml) describes to the ai8x synthesizer how the layers of the neural network should be mapped into the MAX78000 hardware.
Below is how to calculate the number of processors needed in ai85-autoencoder-ecg.yaml
MAX_PROC = 64
def calculate_processors(channels_in: int) -> Tuple[int, int]:
"""
Given a channel count, return the multi-pass adjusted processor count
"""
multipass: int = 1
processors = channels_in
if processors > MAX_PROC: # Multi-pass processor count
multipass = (processors + MAX_PROC - 1) // MAX_PROC
processors = (processors + multipass - 1) // multipass # Rounded up
remainder: int = processors % 4
if remainder != 0:
remainder = 4 - remainder
processors += remainder # To next multiple of 4
return processors, multipassA known answer test (KAT) is also performed on the MAX78000 EVKit. The C code for this KAT can be generated using the scripts/gen_autoencoder_ecg_max78000.sh script in the synthesis repository.
Below is the result from MAX78000 EVKit
---- Opened the serial port /dev/ttyACM0 ----
Waiting...
Measuring system base (idle) power...
*** CNN Inference Test autoencoder_ecg ***
Measuring weight loading...
Measuring input loading...
Measuring input load + inference...
*** PASS ***
See monitor display for inference energy.