Why does the Sigmoid activation function work, but ReLU does not?

There is such a code:

from keras.models import Sequential
from keras.layers import Dense
import numpy
from numpy import exp, array, random, dot

# задаем для воспроизводимости результатов
numpy.random.seed(2)

# разбиваем датасет на матрицу параметров (X) и вектор целевой переменной (Y)
X = array([[0, 0, 1], [1, 1, 1], [1, 0, 1], [0, 1, 1]])

Y = array([[0, 1, 1, 0]]).T

# создаем модели, добавляем слои один за другим
model = Sequential()
model.add(Dense(1, input_dim=3, activation='relu')) # входной слой требует задать input_dim


# компилируем модель, используем градиентный спуск adam
model.compile(loss="binary_crossentropy", optimizer="adam", metrics=['accuracy'])

# обучаем нейронную сеть
model.fit(X, Y, epochs = 10000, batch_size=4)

# оцениваем результат
scores = model.evaluate(X, Y)
print("\n%s: %.2f%%" % (model.metrics_names[1], scores[1]*100))
pre = model.predict(array([[0, 1, 0]]))
print(pre)

precl = model.predict_classes(array([[1,0,0]]))
print(precl)

And when I use the activation function ReLU, the accuracy never increases above 50% and the answers are wrong. When I use Sigmoid, everything converges immediately.

I read that ReLU is now preferred in most cases. What is wrong in my case?

Results (should be: 0, 1):

For ReLU:

accuracy: 50.00%
[[0.9953842]]
[[0]]

For Sigmoid:

accuracy: 100.00%
[[0.13338517]]
[[1]]