Electronic Thesis and Dissertation Repository

Thesis Format



Master of Engineering Science


Electrical and Computer Engineering

Collaborative Specialization

Planetary Science and Exploration


McIsaac, Kenneth


The Canadarm3 is going to operate autonomously aboard the Lunar Gateway space station for the purpose of inspections and repairs. To make the repairs, damage to the spacecraft needs to be detected accurately and automatically. This research investigates methods for training Machine Learning models on 3D point clouds to identify anomalous structural damage. The PointNet algorithm was used to train models on point clouds without affecting their structure. The optimal training data style was found by comparing how well the different styles of data performed at classifying the point cloud testing data. Two different methods of anomaly detection were tested and compared; statistical anomaly detection based on classification scores and anomaly detection using an autoencoder. The autoencoder method proved superior and achieved a recall score of 90.42% with a specificity of 79.31% and a classification score of 97.93%. This showed the potential to use an autoencoder on 3D point clouds for anomalous damage detection on the exterior of spacecrafts.

Summary for Lay Audience

The robotic arm Canadarm3 is being built by Canadian company MDA to perform repairs while operating on the Lunar Gateway space station. Since the arm is being built for both remote and autonomous operation, there needs to be a way to automatically identify damaged parts of the spacecraft exterior so that it will know what parts need to be repaired. Scans of the exterior of the Gateway will be taken using a sensor on the end of the Canadarm3 that produces point clouds, which are 3D representations of an object. This research investigates potential Machine Learning models and methods for training them on 3D point clouds for the purpose of automatically identifying anomalous structural damage.

Various methods for applying deep learning models on point cloud data were researched, and the most promising model architectures were identified and selected for further investigation. To perform the investigation, a procedure for generating training and testing data, both ‘normal’ and ‘damaged’, was developed. The data were generated for simple geometric objects as representations of components which make up the Gateway, with the conjecture that if the models and training would not work on simple objects, they would not work on the more complex spacecraft structures. Multiple styles of training data were tested and compared to each other to determine which style of data provided the best model results.

With the optimal training data style determined, two models were then tested and compared: a classification only model and a multi-output model. The classification only model classified and labeled each point cloud of a simple object, representing a spacecraft component. To evaluate its performance, a statistical method was used to predict if the point cloud contained damage based upon how confident the model was in the label it assigned to the object. The multi-output model had one output which classified the object and another output that reconstructed the point cloud. The reconstructed point cloud was compared against the original point cloud and the difference between them was the reconstruction error. Point clouds with larger reconstruction errors contained damage. The results from each model were compared and the model that achieved the best results was then tuned using hyperparameter tuning to create the best model possible. The results from this optimal model were then analyzed to assess it as a potential solution for automatically detecting real spacecraft damage.