5ARIP10

Predictive Maintentance in Harmonic Drives

Before you get started

To get started, an environment must be created using the environment.yml file. Moreover, to get a full understanding of what the code in this repository does, and what it is used for, the paper 'Predictive Maintenance in Harmonic Drive Systems' should be read.

The repository contains 3 main sections, namely 1. FFT, 2. GA, and 3. NN. To run the code for these three sections, their respective "main..." notebooks can be used. For a more elaborate explanation of the code, please consult the corresponding chapters in this README.

How to run Fast Fourier Transform (FFT)

Relevant files

main_FFT.ipynb

Steps to run

1. Import the packages that are stated in the first cell of the notebook
2. Make a directory FFT_figures_in_eps in your main folder, this is for storing the FFT figures as .eps files
3. Run all cells of the notebook at once

How to run Genetic Algorithm (GA)

Relevant files

main_GA.ipynb
GA_feature_extraction.py

Steps to run

1. Open main_GA.ipynb
2. Run modules "Import libaries" and "Load in the data". This step would generate a database of test data first, and will then load in the microphone and current test data.
   Things to be adjusted:

# adjust the data directory to the one on your own PC please
HD_dl = dl.HD(r'E:\5ARIP0_Interdisciplinary_team_project\data_analyse\5ARIP10\HD_Model_NEW\HD_data')

3. Run module "Extract features". This step will extract features of test data, and form a feature map of the features.
   Things to be adjusted:

# adjust the x's in the below code to change the lengths of time windows during feature extraction

mic_time_feats = fe.extract_time_features(data = mic_data, x = 44100)   # use fs as x
mic_fft_feats = fe.extract_fft_features(data=mic_data, x = 44100)   # use fs as x
# using x = (current_min_len/60) could produce the same n_time_windows as mic data
current_time_feats = fe.extract_time_features(data=current_data, x = int(current_min_len/60))

4. Run module "Perform GA". This step would select the fittest features, and draw a plot between normal and anomalous data with the features (if the number of selected features is set to 2).
   Things to be adjusted:

# choose featuers to be put into genetic_algorithm 
# change the columns of feature_map_norm to choose different features: [:15]->Mic(time); [15:20]->Mic(freq); [20:]->Cur(time)
feature_map_chosen = feature_map_norm[:, :15]

# n_goal_features, num_genes need to be adjusted; population_size, mutation_rate, elitism_rate, num_generations could be tuned. After trials, population_size should increase with num_genes increasing, while other parameters are less likely to be tuned
n_goal_features = 2
data = feature_map_chosen
# if want to cover all combinations of features, then set 2 to the n_genes th, note that the running tiem would be very long 
population_size =  10000  
num_genes = 15  
mutation_rate=0.01
elitism_rate=0.1
num_generations= 5

features_chosen = genetic_algorithm(data, n_goal_features, population_size, num_genes, calculate_fitness, mutation_rate, 
                                    elitism_rate, num_generations)

# note that number of normal test files and anomalous test files should be the same to run this line. Or, adjust n_norm and run.
features_plot(features, n_norm = int(features_chosen.shape[0]/2))

How to run Neural Network (NN)

Relevant files

main_NN.ipynb NN_config.py NN_model.py NN_train.py

Steps to run

1. Open main_NN.ipynb
2. Import the packages
3. Choose whether you would like to use the first itterations procedure, or the second itterations procedure, run the database/dataset generator accordingly.
4. Run the NN creator, a summary of the model is given as output.
5. Run the model trainer
6. Run the plotter, which plots the losses against the epochs. The number or epochs can be determined by selecting an appropriate value from the plot.
7. Run the model tester and find its results in the output.