| Ultrasonic phased array technology is an important detection method in the field of NDT.With the expansion of infrastructure construction in our country,the demand for ultrasonic phased array equipment is growing.The use of TFM in ultrasonic phased array systems can improve the detection accuracy and imaging quality.However,while the traditional TFM is implemented in the time domain,its computational requirement is heavy.So it is difficult to perform real-time detection,and its application in industrial sites is limited.In this paper,the TFM in frequency domain is studied,simulated by MATLAB,and the algorithm is accelerated in parallel with CUDA.Finally,an imaging experiment is carried out.This paper establishes the frequency-domain TFM equation through the study of ultrasonic transmission characteristics,and derives the calculation formulas of FD-SAFT algorithm and FD-FMC.This paper introduces Chirp-Z transform into the TFM in frequency domain,which reduces the fence effect caused by FFT transform and improves the accuracy of Stolt migration.Based on the Field II,the correctness of the frequency-domain TFM is verified.The imaging simulation results show that the frequency-domain TFM has higher computational efficiency than the time-domain algorithm.FD-SAFT has better imaging quality than TD-SAFT.FD-FMC is slightly lower than TD-FMC in imaging quality.In addition,the defects obtained by the frequency domain algorithm are closer to the actual shape,and will not cause diffusion as the depth increases.The paper proposes the design of the frequency domain TFM based on CUDA parallel computing.After giving the overall framework of the algorithm,the implementation of the key modules is demonstrated and compared.The FFT transformation is based on the cu FFT calculation library.The Chirp-Z transformation is completed by converting the time domain convolution operation to the frequency domain multiplication operation.Stolt migration is efficiently implemented by using the hardware interpolation mechanism provided by CUDA texture memory.The vector summation and the maximal value use the parallel reduction algorithm of the interleaved pairing method,and the optimization of shared memory and loop unrolling is carried out.Finally,the whole algorithm is based on CUDA streaming to realize the delay hidden optimization of data transmission between the host and the device.After performance testing,under the author’s experimental platform,the frequency domain TFM design based on CUDA parallel computing can achieve 50 x to 150 x better performance than CPU computing.The paper designs the test software based on TFM in the frequency domain,and uses the self-developed TFM board to conduct the imaging test.The influence of the sampling frequency of the system on the imaging quality is discussed,and the conclusion is that the sampling frequency needs to satisfy the Nyquist sampling theorem of the frequency characteristics of the probe to ensure the imaging quality.Appropriately reducing the sampling frequency under this condition can reduce the computational complexity of the algorithm and the data transmission volume of the system.The experimental results show that FD-SAFT and FD-FMC can show defects clearly with detailed presentation,and the imaging result is in line with expectations.Under the imaging size of 1024×1024,it takes only 1.70 ms for a high-performance workstation equipped with RTX 3060 to complete a single-frame FD-FMC calculation,and a portable notebook equipped with GTX 1050 Ti needs 6.17 ms to complete a single-frame FD-FMC calculation.NVIDIA’s Jetson Nano takes 11.9ms to calculate a single frame of FD-SAFT.In different application scenarios,by flexibly selecting GPUs with different power consumption and computing power,it can meet the needs of real-time TFM. |
PDF Full Text Request