Krishnan Balasubramanian

In this paper, we propose the study of optimizing the hyperparameters of deep learning Data-driven simulation-assisted-Physics learned AI (DPAI) model to simulate the ultrasonic wave propagation for extended depth with a lower error. DPAI model has layers of encoder–decoder structure with modified convolutional long short-term memory (ConvLSTM). DPAI model is trained using the finite element (FE) simulations dataset of distributed single-point to multi-point excitation sources in the 2D domain. The DPAI is the data-driven approach to apprehending the underlying physics of elastodynamic wave propagation. Six different combinations of hyperparameters (hidden dimensions, kernel size, batch size) are used in the DAPI model to study parameter optimization for lowering compounding error. The effectiveness of the trained DPAI models with varying hyperparameters is demonstrated to reduce the compounding error for modeling the deeper simulations of the single-point excitation and multi-point excitation sources. The maximum MAE on amplitude is 5.0×10-2, and MAPE is 2.64% on time of flight (TOF) between DPAI and FE simulations.

In this paper, we proposed an advanced ultrasound imaging technique for enhancing the resolution and reducing the imaging processing time by phased array excitement to create beamforming at virtual source positions. In the conventional full matrix capture (FMC) technique, a single element is used for excitation, limiting the transmitted energy and hence reducing the single-to-noise Ratio (SNR) of the received A-scans. The total focusing method (TFM) is implemented on the vast data volume of the received FMC signals to generate virtually focused imaging, requiring high processing time. Therefore, we have introduced two different scanning techniques to increase transmitted energy by exciting the group of the element to form the virtual source below the transducer: (1) virtual array source aperture (VASA) method consists of multiple virtual sources placed below the center of an active aperture with a fixed focal distance. (2) arbitrary virtual array source aperture (AVASA) method consists of multiple virtual sources located randomly below the active aperture. The ultrasound beam is sequentially excited on each virtual source with predefined delay law. While reception, all the transducers are used to form the FMC. The image generation process is similar to the FMC–TFM method. To demonstrate the imaging capabilities of the proposed techniques, we have performed experimentation on two sets of defective specimens with (a) side-drilled holes (SDHs) and (b) cracks. Experimental results quantitatively compared with conventional FMC–TFM, the proposed method improves the SNR by 35% and reduces computation time by 8 times.

In this paper, we propose a new phased array imaging technique called Arbitrary Virtual Array Source Aperture (AVASA) to image deeper defects with an improved SNR with fewer transmissions. The approach is to transmit the ultrasound waves by electronic beamforming at several arbitrary virtual source positions to achieve higher focal depth to increase the SNR of the received A-scans. Backscattered signals are recorded with all the array elements. A high-resolution image is obtained on reception by virtually focusing on every point in the region of interest by signal coherence summation. In this paper, the proposed AVASA and TFM methods are employed for scanning the larger thickness structure with an unknown defect nature to contrast the defect SNR and the number of defect imaging. Compared with TFM imaging, the AVASA method shows a significantly increasing defect-detecting range with higher amplitude. To further improve the imaging quality and reduce the reconstruction time, the influence of the virtual source parameters on the AVASA imaging and a scanning strategy is demonstrated. A good agreement between the AVASA and TFM is observed, and the number of transmissions is required to inspect the test specimen using AVASA reduced by a factor of four to eight.

In this paper, we propose a framework to automate the process of defect characterizing for industrial structural component health monitoring by implementing automatic defect recognition (ADR) system. The ADR system consists of a convolutional neural network (CNN) and an edge detection algorithm medial axis transform (MAT). The CNN learns the defect feature space from the training dataset to detect and classify the defect. The MAT algorithm is used upon post-validation of the ADR, and the predicted feature’s edges are extracted to size them. The ADR is trained using the simulation-assisted finite element (FE) simulation datasets consisting of side drilled holes (SDH) and crack defects images. The training datasets are generated by introducing virtual array source aperture (VASA), which is a full matrix capture (FMC) scanning strategy by activating the group of elements in an active aperture with predefined focal laws to form a focused beam at a virtual source in the material. The VASA technique uses multiple virtual sources and active aperture positions in a given transducer, which are determined using the Poisson point process. The ultrasound beam is excited in sequence on each virtual source, and the reflected wave is recoded using all the transducers in the array to create FMC A-scans signals. The total focusing method (TFM) technique is a postprocessing algorithm implemented on the FMC signal to generate an image. A large quantity of training datasets is created for each defect by modeling various FE models with varying defect morphology. To create nearly close to experimental images, the experimental noise is introduced in the simulated images. The three separate ADR systems are trained with individual defects class and combined defects. The effectiveness of the trained ADR system is validated by conducting experiments on the plates with laboratory-made SDH and crack defects, the casting components, and weldments with unknown defect types and sizes. The mAP of ADR training is 82%, and the F1-score on testing image classification is 89%. The ADR system could detect and size the smallest defect is 0.219 mm, which is λ L /5.

In this paper, we propose a deep neural network model to simulate the transient ultrasonic wave propagation in the 2D domain by implementing the Data driven-simulation-assisted-Physics learned AI (DPAI) model. The DPAI model consists of modified convolutional long short-term memory (ConvLSTM) with an encoder–decoder structure, which learns the representation of spatio-temporal dependence from input sequence data. The DPAI uses the data-driven approach to understand the underlying physics of elastic wave propagation in a medium. This model is trained with simulation-assisted finite element simulation datasets consisting of distributed single and multi-point excitation sources in the medium. The effectiveness of the proposed approach is demonstrated by modeling a wide range of scenarios in elastodynamic physics, such as multiple point sources, varying excitation parameters, and wave propagation in a large 2D domain. The trained DPAI model is tested and compared against FE modeling with respect to accuracy and computational time.

In this paper, we proposed a data-driven spatio-temporal deep learning (SDL) model, to simulate forward and reflected ultrasonic wave propagation in the 2D geometrical domain, by implementing the convolutional long short-term memory (ConvLSTM) algorithm. The SDL model learns underlying wave physics from the spatio-temporal datasets. Two different SDL models are trained, with the following time-domain finite element (FE) simulation datasets, by applying: (1) multi-point excitation sources inside the domain and (2) single-point excitation sources on the edge of the different geometrical domains. The proposed SDL models simulate ultrasonic wave dynamics, for the forward ultrasonic wave propagation in the different geometrical domains and reflected wave propagation phenomenon, from the geometrical boundaries such as curved, T-shaped, triangular, and rectangular domains, with varying frequencies and cycles. The SDL is a reliable model, which generates simulations faster than the conventional finite element solvers.

In this paper, Artificial Intelligence (AI) algorithms are employed for first, automating the process of creating a large synthetic Total Focusing Method (TFM) imaging dataset using a small set of Finite Element (FE) simulation datasets, and second for the automated defect-recognition (ADR) in butt-welds. In this paper, six types of imaging datasets are created with three approaches. In the first approach, weld TFM images are constructed using ultrasonic A-scan signals obtained from Full Matrix Capture (FMC) performed using FE analysis on models with weld defects (porosity and slag). The second approach generates near real-time weld TFM images by implementing fast deep convolution generative adversarial networks (DCGAN). This second technique permits simulations that are several orders faster when compared to the FE method. In the third approach, noise is extracted from FMC-TFM experimental measurements using the sliding kernel approach, and this noise is supplemented to individual simulated datasets for creating near to realistic scenarios. The first dataset is created using the first approach. The second dataset is created using the second approach, and the third hybrid dataset is a combination of FE and DCGAN weld TFM imaging. The fourth dataset is noise supplemented to FE based dataset. The fifth dataset is generated by adding noise to DCGAN images. The sixth hybrid dataset with noise is a combination of FE and DCGAN weld TFM noise images. AI plays a significant role in object detection and classification through robust feature extraction, reducing human intervention. In this work, for automated weld defect recognition, a convolutional neural network (CNN) is trained using six types of simulation-assisted weld TFM imaging datasets, which improves the reliability and efficiency of welds quality assurance. The mAP value is 85% for the ADR model trained using the hybrid weld TFM dataset with noise. The model prediction on classification on the hybrid dataset for porosity is 0.86 F1-score, and for slag is 0.80 F1-score.