Test_helm.py

"""
This file is subject to the terms and conditions defined in
file 'LICENSE.txt', which is part of this source code package.

Code for training and running the HELM as in
Michau, Gabriel, Yang Hu, Thomas Palmé, and Olga Fink. “Feature Learning for Fault Detection in High-Dimensional Condition-Monitoring Signals.”
ArXiv Preprint ArXiv:1810.05550, 2018.

Author: Gabriel Michau

Version: 30/10/2018
"""

import numpy as np
import random
import matplotlib.pyplot as plt

from lib import mapminmax as mmm
from lib import HELM

# %%
train_x = np.random.rand(6000, 200)
vali_x = np.random.rand(1000, 200)
test_x = np.random.rand(2000, 200)
fault_id = random.sample(range(200), 5)
test_x[1001:2000, fault_id] = test_x[1001:2000, fault_id] + 2

train_x, paramap = mmm.mapminmax(train_x, q=0, axis=1, mini=-1, maxi=1)
vali_x, test_x = mmm.mapminmax_a(paramap, vali_x, test_x)
# Hyperparameters of HELM
para = {'nhelm': 5, 'neuron_number': np.array([20, 100]), 'fista_weight': 1e-3, 'fista_cv': 1e-5, 'ridge_weight': 1e-5}

# %% Train HELM
model = HELM.HELM(para, train_x)
# %% Run HELM on the three datasets
out = HELM.HELM_run(model, train=train_x, val=vali_x, test=test_x)

# %% Plot results

# Define a detection threshold based on the quantile of the validation set
quant = 99.9
thr = 1.5 * np.percentile(out['val']['Y'], quant)

f1: object = plt.figure(1)
f1.clf()
plt.plot(range(6000), out['train']['Y'] / thr, linestyle='None', marker='o', markerfacecolor='none')
plt.plot(range(6000, 7000), out['val']['Y'] / thr, linestyle='None', marker='o', markerfacecolor='none')
plt.plot(range(7000, 9000), out['test']['Y'] / thr, linestyle='None', marker='o', markerfacecolor='none')
plt.plot([0, 9000], [1, 1], color='k')
plt.show()