LSTM time series data classification model low accuracy - keras

Question

I'm trying to create a model for time series analysis using LSTM layer, however accuracy is very low even when using Dense layers and no LSTM.
The data is time series (synthetic spectrum), which depends on 4 parameters. Changing the parameters enables to use different size datasets, where each sample is more or less different from the other. But no matter the size of dataset accuracy is always as low as 0.0 - 0.32 %.
Model with LSTM:
print(trainset.shape)
print(testset.shape)
print(trainlabels.shape)
model = Sequential()
model.add(Masking(mask_value=0.0, input_shape=(trainset.shape[1], trainset.shape[2])))
model.add(LSTM(10, activation='relu', stateful=False, return_sequences=False))
model.add(Dropout(0.3))
model.add(Dense(len(trainlabels), activation='relu'))
model.compile(loss='sparse_categorical_crossentropy',
optimizer='Adam', metrics=['accuracy'])
print(model.summary())
model.fit(trainset, trainlabels, validation_data=(testset, testlabels),
epochs=3, batch_size=10)
scores = model.evaluate(testset, testlabels, verbose=0)
print("Accuracy: %.2f%%" % (scores[1]*100))
OUTPUT:
(2478, 600, 1)
(620, 600, 1)
(2478,)
Model: "sequential_7"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
masking_7 (Masking) (None, 600, 1) 0
_________________________________________________________________
lstm_7 (LSTM) (None, 10) 480
_________________________________________________________________
dropout_7 (Dropout) (None, 10) 0
_________________________________________________________________
dense_7 (Dense) (None, 2478) 27258
=================================================================
Total params: 27,738
Trainable params: 27,738
Non-trainable params: 0
_________________________________________________________________
None
Train on 2478 samples, validate on 620 samples
Epoch 1/3
2478/2478 [==============================] - 53s 22ms/step - loss: 8.9022 - accuracy: 4.0355e-04 - val_loss: 7.8152 - val_accuracy: 0.0016
Epoch 2/3
2478/2478 [==============================] - 54s 22ms/step - loss: 7.8152 - accuracy: 4.0355e-04 - val_loss: 7.8152 - val_accuracy: 0.0016
Epoch 3/3
2478/2478 [==============================] - 53s 21ms/step - loss: 7.8152 - accuracy: 4.0355e-04 - val_loss: 7.8152 - val_accuracy: 0.0016
Accuracy: 0.16%
Some values in training data are 0.0, therefore Masking is used.
I have tried playing around with different loss, optimizer, activations, Dropout and layer parameters. The result is always the same even after adding more Dense layers, or changing batch size.
Model with Dense:
Data format is 2D instead of 3D.
model=keras.Sequential([
keras.layers.Masking(mask_value=0.0, input_shape=(trainset.shape[1],)),
keras.layers.Dense(128, activation='relu'),
keras.layers.Dropout(0.05),
keras.layers.Dense(64, activation='relu'),
keras.layers.Dropout(0.05),
keras.layers.Dense(len(labels), activation='softmax')
])
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.fit(np.uint8(trainset), np.uint8(trainlabels), epochs=100)
test_loss, test_acc=model.evaluate(np.uint8(testset), np.uint8(testlabels),
verbose=2)
print(test_acc)
OUTPUT:
Train on 1239 samples
Epoch 1/100
1239/1239 [==============================] - 1s 1ms/sample - loss: 8.5421 - accuracy: 0.0033
Epoch 2/100
1239/1239 [==============================] - 0s 371us/sample - loss: 6.2039 - accuracy: 0.0025
Epoch 3/100
1239/1239 [==============================] - 0s 347us/sample - loss: 5.6502 - accuracy: 0.0033
****
Epoch 97/100
1239/1239 [==============================] - 0s 380us/sample - loss: 0.1472 - accuracy: 0.9746
Epoch 98/100
1239/1239 [==============================] - 0s 364us/sample - loss: 0.1562 - accuracy: 0.9680
Epoch 99/100
1239/1239 [==============================] - 1s 408us/sample - loss: 0.1511 - accuracy: 0.9721
Epoch 100/100
1239/1239 [==============================] - 0s 378us/sample - loss: 0.1719 - accuracy: 0.9680
310/1 - 0s - loss: 18.6845 - accuracy: 0.0000e+00
0.0
With this model loss is very low but so is accuracy.
What kind of model architecture should be used for my data?
Thanks in advance for helping to learn this stuff!