The project investigates a signal represented as a sum of harmonic functions with added noise. For convenience, the signal is discretized into a set of 1000 points. A peak is defined as a point where the value significantly exceeds those of neighboring points. Experiments show that the scipy.signal.find_peaks function does not always accurately classify peaks, so a fully connected neural network is used for peak detection, as was proposed by V. D. Neverov.
peak_detection.ipynb— Jupyter notebook containing code for generating signals, building the dataset, creating and training the neural network, and visualising peaks.- signals_dataset.npz — pre-generated dataset of synthetic signals and corresponding peak masks.
- signal_model.keras — saved neural network model.
- signal_model_weights.h5 — saved weights of the neural network.
- Generates synthetic signals of length 1000 with multiple sinusoidal components and random noise.
- Constructs a mask array indicating positions of peaks.
- Saves the dataset in
.npzformat for reuse without regeneration. - Implements a fully connected neural network that outputs a continuous array of length 1000 with values from 0 to 1 representing peak probabilities.
- Uses MSE as the training and evaluation metric.
- Provides visualization of true and predicted peaks.
The model was trained for 60 epochs and did not overfit, achieving a final MSE of 0.002730 for signals of 1000 points with an average of 17.5 peaks.