---
library_name: keras
tags:
- timeseries
---
# Timeseries classification from scratch
Based on the _Timeseries classification from scratch_ example on [keras.io](https://keras.io/examples/timeseries/timeseries_classification_from_scratch/) created by [hfawaz](https://github.com/hfawaz/).
## Model description
The model is a Fully Convolutional Neural Network originally proposed in [this paper](https://arxiv.org/abs/1611.06455).
The implementation is based on the TF 2 version provided [here](https://github.com/hfawaz/dl-4-tsc/).
The hyperparameters (kernel_size, filters, the usage of BatchNorm) were found via random search using [KerasTuner](https://github.com/keras-team/keras-tuner).
## Intended uses & limitations
Given a time series of 500 samples, the goal is to automatically detect the presence of a specific issue with the engine.
The data used to train the model was already _z-normalized_: each timeseries sample has a mean equal to zero and a standard deviation equal to one.
## Training and evaluation data
The dataset used here is called [FordA](http://www.j-wichard.de/publications/FordPaper.pdf). The data comes from the [UCR archive](https://www.cs.ucr.edu/%7Eeamonn/time_series_data_2018/). The dataset contains:
- 3601 training instances
- 1320 testing instances
Each timeseries corresponds to a measurement of engine noise captured by a motor sensor.
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
| name | learning_rate | decay | beta_1 | beta_2 | epsilon | amsgrad | training_precision |
|----|-------------|-----|------|------|-------|-------|------------------|
|Adam|9.999999747378752e-05|0.0|0.8999999761581421|0.9990000128746033|1e-07|False|float32|
## Model Plot
View Model Plot
Model reproduced by Edoardo Abati