Abstract: Due to the significant variations in size and morphology among different types of wear particles, and the limited depth of field of microscopes, particles of varying thickness may appear defocused and blurred within a single ferrograph image. To address the challenges of omission and misidentification caused by defocused particles in single-image analysis, a progressive layer-by-layer recognition algorithm for wear particles in ferrograph sequence images was proposed. First, an instance segmentation model, termed WearIS, was developed for a single ferrograph image. This model incorporated the Convolutional Block Attention Module (CBAM), a variance-cascaded head network, and a segmentation branch fusing deep and shallow features to accurately identify clear wear particles in the images. Second, a progressive layer-by-layer recognition algorithm was designed to iteratively refine the recognition results across sequence images, which utilized metrics such as the intersection-over-union (IoU) of overlapping particles between consecutive frames and confidence scores. This algorithm performed frame-by-frame association and correction, ultimately ensuring comprehensive identification of all wear particles within the sequence. Comparative experimental results demonstrated that the proposed algorithm achieved detection and segmentation AP50 values of 82.67% and 80.92%, respectively, and a mean IoU of 75.64% on the ferrograph sequence image test set, with an average processing time of 1.07 seconds per frame. Compared to single-image ferrograph analysis methods, the proposed approach significantly enhanced wear particle recognition accuracy while effectively reducing the probability of omission or misidentification of anomalous particles.