Battery defect detection using ultrasonic guided waves and a convolutional neural network model

Energy storage batteries play a crucial role in regulating modern power grids. However, energy storage systems face numerous safety risks, with battery safety being the primary constraint on their widespread application. To ensure the safety of these batteries during operation, this paper presents a battery defect detection method based on ultrasonic guided waves and a Convolutional Long Short-Term Memory (ConvLSTM) neural network model. By utilizing lead zirconate titanate (PZT) ultrasonic transducers to excite ultrasonic guided waves within the battery and employing a laser Doppler vibrometer (LDV) to receive ultrasonic signals from the battery surface, the entire propagation process of the ultrasonic guided waves within the object is captured. To enhance the robustness of the model, data augmentation techniques are applied. By augmenting the data from pouch battery test samples, the number of effective images for model training was increased from 3400 to 15,300. To learn from the dataset of defective batteries and prevent gradient explosion during training, this paper proposes an algorithm inspired by the U-Net model, which incorporates spatiotemporal information. The algorithm introduces a ConvLSTM module, enabling the model to learn and retain temporal features from time-series images, and to segment and identify internal battery defects based on the propagation characteristics of ultrasonic guided waves. Experimental comparisons demonstrate that the proposed method excels in identifying internal defects in batteries, with an intersection over union (IoU) of 80.32 % and a pixel accuracy (PA) of 96.47 %. This method provides a convenient and accurate implementation scheme for the safety testing of energy storage batteries or the safe appraisal of retired batteries.