XNN

Explainable Neural Networks for Multivariate Time Series

Article
Github
Contact

Save model load model

epochs

batch

Input
x-axis: feature sequences
y-axis: engines

Output
x-axis:
y-axis: engines

Filter weight (absolute)

Close

Please select a load source

Select a model from server

Select a model from browser local storage

Upload files

Cancel

Model Name

Local storage

Downloads

Save & close Cancel

Select & close Cancel

Layer

Units

Activation

OK Cancel

The article

Quick start:

Click the start/pause button to start/pause the training process accordingly. The training loss vs. testing loss graph will be updated every batch. The intermediate outputs, final training output, final testing output, and weights are updated every epoch.

Functionalities

Model configuration and training
Click on button to delete the layer.
Click on button to add more layer.
Change number of epochs/batch size if needed.
Click on snapshots checkbox to enable/disable saving training snapshots, so user could review the trained model later on.
Click on button to start training process.
Model weights and intermediate outputs view
Click on the weight type (e.g., LSTM, positive/negative) to toggle display.
Mouse over the weight path to see its current value.
Load model
Click on load model button to load and review the a saved, trained model.

Data Description

The simulation model is Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset contains 27 attributes of numerical values. Each row is a record of an operational cycle of an individual unit. There are three operational settings called Altitude, Mach number, and Throttle Resolver Angle (TRA). These parameters have a substantial effect on engine performance. The next 21 columns show the sensor data, which has been recorded from each cycle and is contaminated with sensor noise. There are a training set and a testing set for each dataset. In the training set, the engine ran to failure. In the test set, the time series ended sometime before the engine fully degraded or failed, and the prognostics algorithm was tasked to predict the RUL.

Graph description

Please click on each of the graphs to view the graph in details.

After loading, the system displays the input in the form of the heat-maps, one heat-map per sensor data column. The system does not display the axes for graphs at the input layer. Instead, it gives their descriptions on top of each layer for the simplicity of the visual display. The x-axis of the heat-map is the engine units (e.g., 0-99 for dataset 1). Its y-axis is the sensor data sequence (the number of cycles). The heat-map color at a specific point represents the sensor's value of a corresponding engine at a time step accordingly. The input units are also sorted by their RUL to see if they appear to have any sensor input pattern. The first panel of the picture above shows the details view of a sample input for one sensor.

Similar to the input layer, the outputs from LSTM layers are also sequences with the corresponding number of engines and steps. Therefore, the system also represents them as heat-map (second panel) with a similar specification. The second panel of the picture above shows the details view of a sample LSTM layer output.

Outputs of the Dense layers, and the final training and testing outputs are alike and are visualized as scatter plots. Y-axis also represents the number of engines, and x-axis describes the output values. The green circles at the scatter plot represent the predicted outputs of the corresponding engine units as inputs. The gray x symbols are the actual target RULs. The two are scaled linearly to the domain of target RULs for better visualization. The second panel of the picture above shows the details view of a sample Dense layer output.