I am wanting to use an LSTM for anomaly detection on a multivariate time series data. Let's say there are n rows each corresponding to a timestamp incrementing by an hour and d input features and d output flags for each feature. These flags are 0,1 or 2. I know how to shape the input data (samples, timesteps, features. So If I take m timesteps in a sample, when I give a 2D input as (1,m,d), I want the output to be the matrix m*d (which is the d output flags for each timestep across m timesteps in the sample.

  1. How do I shape the output data and layers of my NN for this ?

I have tried using keras.preprocessing.timeseries_dataset_from_array. However, for an input of (batch_size, timesteps, input features) it gives as output data (batch_size, output features) instead of (batchsize, timesteps, outpupt features) I would be glad to give additional information as needed and would greatly appreciate any help. Thanks and regards



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.