Dimensional Prediction And Surface Morphology Inspection Of Deposited Layers For Laser Deposition Additive Manufacturing Based On Machine Learning Principles

Directed energy deposition(DED)is a form of additive manufacturing(AM)and is widely used as a novel modern manufacturing technique where parts are manufactured layer by layer using digital models.This method of additive manufacturing has the advantages of fast and efficient forming and the ability to produce large components.However,there are still many issues that affect the additive manufacturing process and the performance of the part.It is well known that the shape of the melt pool is influenced by the process parameters during the additive manufacturing process,which can affect the performance of the part.In addition,the shape of the deposited layer during the forming process also influences the subsequent performance of the overall component.It is therefore of great interest to investigate the relationship between melt pool dimensions and process parameters as well as to inspect the surface shape of the deposited layer in real time.Therefore,this paper conducts a study on the prediction of deposition layer dimensions based on machine learning methods.Three machine learning models,namely support vector regression(SVR),extreme gradient boosting(XGBoost)and back propagation neural network(BPNN),are experimentally selected and constructed to predict the melt pool shape and maximize the prediction performance by integrating various kernel functions,optimization algorithms and model parameters.The geometry of the deposition layers,such as the height,width and depth of 210 single-channel deposition layers,were also measured as training and test datasets to provide the base data for the prediction models when training and testing.In addition to the training and test datasets,a new dataset with 36 samples was prepared to validate the generalization performance and prediction accuracy of the three machine learning models on the completely new data.The results showed that on the test dataset,the RBF kernel-based SVR model predicted the melt pool height with ～93% accuracy,while the XGboost model achieved ～97% and 96.3% accuracy for the melt pool width and depth,respectively.In the new dataset,the BPNN model based on Adam’s method achieves ～93.7%accuracy for the melt pool forming height,but the XGboost model achieves ～96.6% and 97.8%accuracy for the melt pool forming width and depth.Thus,in addition to the training and test datasets,the optimized machine learning models demonstrated significant prediction accuracy,excellent generalization and robustness in the new dataset.As a result,the application of machine learning has greatly improved the possibility of controlling the laser deposited process more intelligently and stably.Furthermore,innovative applications of machine learning and neural network algorithms in additive manufacturing processes have been extensively studied recently,but the development and practical application of this technology remains problematic.To solve the problem of difficult and costly preparation of surface topography data due to the large number of samples required for visual inspection using a neural network framework.In this study,a transfer learning method based on a small dataset is proposed to detect surface morphology during laser deposition.The proposed method effectively and innovatively solves the problem of sample scarcity that arises during the training of neural networks and eliminates the problem of having to re-prepare a considerable volume of surface topography datasets.It also effectively accelerates the convergence process of the model through a transfer learning approach,indirectly improving the detection accuracy.This work chose a parametric transfer learning approach to train YOLOv7,a self-supervised object detection model,for transfer learning.The experiments utilized an open-source defect dataset as a pre-training dataset for transfer learning pre-training.Formal model training was performed after transferring the model parameters in combination with a self-made target dataset containing 240 small samples.The input to the models was image data and the output of the training was surface morphology classification.The results show that the transfer training model 2 has a high accuracy and check-all rate,achieving an average accuracy value of 0.62.Transfer learning speeds up the feature extraction process by directly transferring the parameters of a pretrained model for a similar task into the current task model,thereby speeding up the training process and improving detection performance.