Study On Intelligent Ultrasonic Testing Methods For Metallurgical Characteristics Of Laser Additively Manufactured Metals

Metal laser additive manufacturing technology can realize the integrated net forming of complex structure metal components,which is a key technology to improve the design and manufacturing capabilities of high-precision fields such as aerospace.The technology uses a laser beam to melt powdered raw materials and accumulate them layer by layer to produce parts with complex shapes and excellent properties.However,due to the complex heat transfer phase change and solid-liquid-gas coupling physical processes involved in metal laser additive manufacturing,it is easy to form coarse columnar crystals with obvious anisotropy and prone to cracks and hole-like defects,which greatly reduce the reliability of the parts.Therefore,it is essential to perform rapid and accurate non-destructive testing of grain size and defects of metal laser additive manufacturing formed parts.Laser ultrasonic technology can realize remote and non-contact scanning inspection of internal organization and defects of materials in the industrial environment,and has great application prospects in online inspection and closed-loop quality control of the manufacturing process.Aiming at the problem of low precision of non-destructive evaluation of grain size in metal laser additive manufacturing caused by insufficient artificial prior knowledge,this thesis proposes an intelligent nondestructive evaluation method of grain size by laser ultrasound based on continuous wavelet transform and convolutional neural network.Firstly,Ti6Al4V/B₄C samples with different grain sizes were prepared by Laser Melting Deposition（LMD）technique by adding different proportions of nanoscale B₄C powder into Ti6Al4V powder.The nano-pulsed solid-state laser was used to excite high-frequency ultrasound,and the echo ultrasonic signal was transmitted to the bottom surface of the opposite side of the sample based on the dual-wave mixed interference principle.Then,the echo signal was preprocessed by"Derivation-Denoising-Normalization",and the Cmor wavelet basis function was selected to carry out the time-frequency analysis of the preprocessed signal by continuous wavelet transform.The optimal center frequency and bandwidth of the wavelet basis function were determined according to the minimum Shannon wavelet entropy,and different combinations of center frequency and bandwidth parameters around the optimal parameters were selected to obtain the wavelet time-frequency atlas,thus the training sample data was expanded.Based on the VGGNet model,the deep convolutional neural network model was constructed by repeatedly stacking the convolutional kernel and the maximum pooling layer.With wavelet time-frequency maps as the input and grain size as the predicted output,the network model was trained by the validation dataset and training dataset,and the model was optimized by changing the network structure and adjusting the hyperparameters.Finally,the evaluation accuracy of the network model was evaluated by the test data set.Experimental results showed that the proposed method can accurately predict the microstructure grain size of Ti6Al4V/B₄C,and the average prediction accuracy reached 98.26%.The quality and evaluation accuracy of online defect imaging in the metal laser additive manufacturing process was limited by the sampling frequency of online inspection equipment,and high-resolution sampling will reduce the efficiency of online inspection and increase the cost of equipment.To solve the above problems,this thesis proposes a laser ultrasound imaging technique for defects based on compressive sensing.Taking the rolled 2021 aluminum alloy and LMD Al Si10Mg aluminum alloy with preset hole defects as the research object,the laser ultrasonic imaging online inspection system was used to realize two-dimensional fast traversal scanning of the sample surface by vibrating mirror excitation pulse laser beam scanning and continuous detection laser fixation.The laser ultrasonic surface wave signal was acquired at a low sampling frequency,and the high-frequency electrical noise was eliminated by averaging the signals obtained by multiple excitations of a single point.Then the compressed sensing technology was used to reconstruct the averaged signal to obtain a high-resolution reconstructed signal.After denoising the reconstructed signal,the defect imaging was carried out by the variable time window energy mapping method.Three greedy compressed sensing algorithms,Matching Pursuit（MP）,Orthogonal Matching Pursuit（OMP）,and Compressive Sampling Matching Pursuit（Co Sa MP）,were used to reconstruct the single point A-scan signal and the defect image respectively.The differences in reconstruction effects were compared according to the Root Mean Square Error（RMSE）,and the OMP algorithm was determined to be the optimal reconstruction algorithm.Furthermore,the key parameters of the reconstruction algorithm,such as sampling rate and sparsity,were optim ized.The results showed that compared with the defect image obtained by high-resolution sampling,the standard deviation of imaging error of the reconstructed defect image obtained by the OMP compressed sensing algorithm at each scanning point was reduced by 6 times compared with the low-resolution sparse sampling defect image,which verifies the effectiveness of the defect reconstruction imaging method based on online detection of sparse sampling signal compressed sensing.