Bolts, functioning as critical components in connecting structures, play an essential role across various engineering industries. The guided wave method has demonstrated significant potential in detecting bolt looseness with high efficiency. However, identifying the looseness condition of the joint with multiple bolts remains challenging. Various machine learning methods were introduced to extract features from the received signals, yet those that possess a clear physical interpretation in engineering tend to achieve superior application outcomes. Therefore, a physics-informed convolutional neural network (PICNN) is presented for bolt looseness localization and severity estimation for a lap joint connected by eight bolts. A small number of SH-typed magnetostrictive transducers, arranged in a pitch-catch configuration, were used to obtained the transmitted waves from various wave propagation paths. The combined time-frequency spectrum derived from the wavelet transform results of four transmitted waves was used as the input to the model. The relationship between the normalized wave energy transmission ratios and the local bolt looseness severity was revealed by the backpropagation neural network, which is regarded as the wave energy propagation mechanism within the PICNN. Numerical and experimental results indicate that the bolt looseness conditions can be successfully estimated. Time masking and frequency masking data augmentation was performed on the limited experimental samples, and the transfer learning technique was proposed to enhance the bolt looseness detection accuracy.