Online defect detection using a high-speed multi-sensor data fusion system for laser metal deposition

Error on title page – year of award is 2022.Laser metal deposition (LMD) is an advanced additive manufacturing technology that is also known as metal 3D printing. It has many industrial applications which include building parts with complex geometry from scratch, Remanufacturing, and coating parts with a variety of materials. It is being adopted by the industry at a very quick rate, however quality assurance is still a hurdle that is being investigated. Defects are an inherent part of the process and are caused by many factors which may be unacceptable to industries like the Aerospace industry. For this the state of the art presents many NDT solutions to detect defects during the deposition process however, these methods have limitations due to the type of sensor being used, the high rate of change of the phenomenon the sensor is observing, the level of information that can be extracted from the sensor data about the defects and obviously the accuracy of the extracted information. This research investigates and develops a defect detection methodology that uses a multisensory array in collaboration with a custom data fusion algorithm that allows for the detection of defects and predicts features of the detected defects. This includes total number of defects, types and quantity of the defect, Max defect size and total defected area. This is achieved by first designing and developing a multisensory architecture capable of monitoring different parts of the defect development cycle at high enough sampling speeds so that information that indicates defects can be effectively captured. This is followed by monitoring defect provocation experiments to capture signals from a defected sample and training the system on these signals. In the training run the system takes the signals from these defect provocation experiments and stiches them onto defect information extracted from the XCT scans from the provocation experiments. From the stitched data sets events that exhibit anomalous behaviour which might indicate the presence of a defect are extracted and plugged into a K means clustering algorithm which sorts them into clusters. For every cluster a predictor table is formed which are used to predict defect features for any new event that is assigned to that cluster. The online data fusion algorithm takes in values from the predictor table once a new defect is introduced and outputs a predicted range between which the actual value of the defect feature lies in. The reliability of the range is also quantified using another value called % confidence which is formed using a unique scoring system. The results of this system show a relatively higher accuracy than solutions that use a single sensor approach and predicts further information about the defects.Laser metal deposition (LMD) is an advanced additive manufacturing technology that is also known as metal 3D printing. It has many industrial applications which include building parts with complex geometry from scratch, Remanufacturing, and coating parts with a variety of materials. It is being adopted by the industry at a very quick rate, however quality assurance is still a hurdle that is being investigated. Defects are an inherent part of the process and are caused by many factors which may be unacceptable to industries like the Aerospace industry. For this the state of the art presents many NDT solutions to detect defects during the deposition process however, these methods have limitations due to the type of sensor being used, the high rate of change of the phenomenon the sensor is observing, the level of information that can be extracted from the sensor data about the defects and obviously the accuracy of the extracted information. This research investigates and develops a defect detection methodology that uses a multisensory array in collaboration with a custom data fusion algorithm that allows for the detection of defects and predicts features of the detected defects. This includes total number of defects, types and quantity of the defect, Max defect size and total defected area. This is achieved by first designing and developing a multisensory architecture capable of monitoring different parts of the defect development cycle at high enough sampling speeds so that information that indicates defects can be effectively captured. This is followed by monitoring defect provocation experiments to capture signals from a defected sample and training the system on these signals. In the training run the system takes the signals from these defect provocation experiments and stiches them onto defect information extracted from the XCT scans from the provocation experiments. From the stitched data sets events that exhibit anomalous behaviour which might indicate the presence of a defect are extracted and plugged into a K means clustering algorithm which sorts them into clusters. For every cluster a predictor table is formed which are used to predict defect features for any new event that is assigned to that cluster. The online data fusion algorithm takes in values from the predictor table once a new defect is introduced and outputs a predicted range between which the actual value of the defect feature lies in. The reliability of the range is also quantified using another value called % confidence which is formed using a unique scoring system. The results of this system show a relatively higher accuracy than solutions that use a single sensor approach and predicts further information about the defects