Non-destructive testing (NDT) of magnetic materials such as aviation parts is an indispensable part of the civil aviation maintenance industry. The NDT of such metal materials often uses magnetic particle inspection (MPI) technology. This paper proposes an improved DeepLabv3+ semantic segmentation algorithm for automatic defect detection of aviation ferromagnetic parts after MPI. In the network structure, lightweight MobileNetV2 is the backbone feature extraction network. The dense atrous spatial pyramid pooling (DenseASPP) structure is used to strengthen feature extraction. The influence of three different DenseASPP structures on the recognition effect is compared in the experiment. At the same time, the decoder is further optimized. The experimental results show that the Ours-DeepLabv3+ network model can effectively for automatic defect detection of aviation ferromagnetic parts after MPI. The Precision, Recall, F1-score, and intersection over union are 81.64%, 83.12%, 82.37%, and 71.23%, respectively, which are 7.48%, 5.45%, 6.50%, and 10.1% higher than the original DeepLabv3+, and defect detail segmentation is more accurate. Compared with other semantic segmentation algorithms, this method can effectively improve the accuracy of defect detection of aviation ferromagnetic parts and meet the requirements of defect detection.

The design of a whisker sensor, inspired by mammalian whisker characteristics, is presented in this paper. It uses a novel spring structure to transfer the deformation generated by the whisker tip when it touches an object at the base, which drives the permanent magnet installed at the base to change its position. It achieves precise positioning of the object by using the magnetic induction intensity data output from the Hall sensor MLX90393. Based on the results of the finite element model analysis, the detection range of the whisker sensor can be expanded by replacing the artificial whisker material and selecting a permanent magnet of a suitable size. Calibration experiments and positioning tests were conducted on the sensor. The experimental results showed that the detection radius of the sensor was 24, 30, 33, and 39 mm for the carbon fiber, acrylic, acrylonitrile butadiene styrene plastic (ABS), and nylon whiskers, respectively, when they were matched with a NdFeB annular permanent magnet with an aperture of 3 mm and a thickness of 3 mm. The sensor is small and simple to manufacture with good sensitivity, linearity, hysteresis, and repeatability. The maximum positioning errors of the X and Y positions in the detection plane of the sensor were within ±1.3 mm, and the positioning was accurate. The sensor can be used to identify the shape of an object.

This paper introduces a novel Hall-effect-based whisker sensor and proposes a contour recognition algorithm for distinguishing and recognizing objects. The mechanical arm drives the sensor to move according to the contour recognition algorithm in the experiments, and the object contour is distinguished and recognized by processing the detected spatial magnetic field data. The results of identifying the measured object with harder carbon fiber and softer ABS whiskers were compared and analyzed, and a new design was proposed based on the advantages of soft and hard materials. The newly designed whisker with glass fiber and stainless steel rod nested reduced the contour recognition error to 4.1%. The experimental results showed that the sensor could distinguish the contour of the straight line and circular objects, and the glass fiber and stainless steel rod nested sensor had a better recognition effect.

Tool wear is inevitable in machining and directly impacts the forming quality of the workpiece. Accurate monitoring of tool wear state can effectively improve machining efficiency and reduce the adverse effects caused by tool wear. This paper proposes an intelligent recognition approach for tool wear state in machining based on vibration signals. The EMDResNeStTime (ERT) module uses empirical mode decomposition (EMD) and two-dimensional convolution to extract the main trend features of vibration signals, and the Sequence-Global-Encoding (SGE) module uses the self-attention mechanism to extract the global features of vibration signals, are designed. A new Multi-feature Parallel-time (MFPT) feature extraction backbone is constructed using these two modules. The use of this backbone effectively improves the ability of the network model to extract key features of complex trends in vibration signals. Experimental results show that the proposed approach achieves 71.65%, 74.89%, 71.36% and 72.78% in accuracy, precision, recall and F1 score, respectively. Compared with other network models, it has higher recognition accuracy and more stable performance. The ablation experiment further demonstrates the classification effectiveness of the method used in this network model.

The safe and reliable operation of aero-engines critically depends on the structural integrity of their key components. To address the challenges in aero-engine defect detection, including diverse morphologies, complex backgrounds, small targets, and large-scale variations, a geometry-adaptive and context-optimized fusion network (GC-Net) was proposed. To enhance feature extraction for geometrically complex defects, an adaptive geometric convolution module dynamically aligns convolution kernels with irregular contours of defects via deformable sampling and an edge-aware mechanism. A cascaded context optimization module was designed to integrate dual-path channel aggregation and dual-axis spatial attention to suppress background interference and improve the feature representation of small targets. A dual-direction detail-enhanced fusion network was constructed to facilitate deep interactions across multi-scale features and improve detection consistency for defects at different scales. A cross-scale fusion detection head was developed to achieve dynamic multi-scale feature fusion and accurate bounding box localization. Furthermore, a composite regression loss function combining scale‐dynamic loss and shape‐aware IoU was designed to further optimize the model’s performance in detecting multi‐scale and irregularly shaped defects. Experimental results on a self-built aero-engine borescope image dataset show that GC-Net surpasses the baseline model by 4.8% in mAP ₅₀ and 3.8% in mAP _50:95 , with notable relative improvements on small and medium-sized defects despite severe class imbalance inherent in real-world inspection data. The model achieves a recall of 81.5% and a precision of 93.3%, while maintaining 5.1 M parameters and 135.5 FPS. Cross-domain experiments on the public industrial surface defect datasets GC10-DET and NEU-DET further validate the robustness and transferability of the proposed method, demonstrating its potential applicability beyond the specific aero-engine domain. GC-Net achieves an excellent balance between accuracy, speed, and model complexity, offering a reliable and efficient visual inspection solution for the intelligent maintenance of aero-engines.