Redundant Parameters Research Articles

Error Sensitivity Analysis of a 1T2R Kinematically Redundant Parallel Mechanism With Closed-Loop Chain

Abstract In this paper, a generalized method for error modeling of the spatial 1T2R three degrees-of-freedom kinematically redundant parallel mechanism with a closed-loop chain is proposed, which is based on the matrix differential method. First, the detailed process of generalized error modeling and error analysis are described. Based on the proposed method, the error model of the spatial 3PRR(RR)S-P (P—prismatic joint, R—revolute joint, S—spherical joint, and the underline indicates that the joint is the actuator) kinematically redundant parallel mechanism is established as an example, and the correctness of the error model is verified by combining forward with inverse kinematics. Then, the patterns affecting the output error of the moving platform are discussed for the case where the mechanism contains only static error or dynamic error, respectively. In addition, the error sensitivity indices are defined to evaluate the error sensitivity of the moving platform to different redundant parameters L4 under a certain pose. Finally, in order to identify the key error terms, the sensitivity of the output error of the mechanism to a single error term is analyzed. The results show that the error sensitivity of the spatial kinematically redundant parallel mechanism can be effectively reduced by adjusting the kinematically redundant parameters, so that the mechanism can maintain a low error sensitivity in a certain pose.

Kinematic calibration of a 5-DOF double-driven parallel mechanism with sub-closed loop on limbs

Abstract This paper proposes a kinematic calibration method of a novel 5-degree-of-freedom double-driven parallel mechanism with the sub-closed loop on limbs. At first, considering the introduction of a sub-closed loop significantly increased the complexity and difficulty of kinematic error modeling, an equivalent transformation method is proposed for the limb with a sub-closed loop. Then kinematic error model of the parallel mechanism is established based on the closed-loop vector method and parasitic motion analysis, which is verified by virtual prototype technology. Because the full kinematic error model is generally redundant, error parameter identifiability analysis is carried out by QR decomposition of the identification Jacobian matrix, and the redundant parameters are removed. Additionally, the Sequence Forward Floating Search algorithm is utilized to optimize measurement configurations to reduce the influence of measurement noise. Finally, with a laser tracker as the measuring device, numerical simulations and experiments are implemented to verify the proposed kinematic calibration method. The experiment results show that average position and orientation errors are reduced from 2.778 mm and 1.115° to 0.263 mm and 0.176°, respectively, within the prescribed workspace.

Facial expression recognition based on convolutional network attention mechanism

Aiming at the problems of large number of parameters and weak recognition ability in expression recognition, a method of facial expression recognition based on convolutional network is proposed. An improved residual module is introduced to reduce the number of parameters while enhancing the attention to the expression area; the channel-space attention mechanism is used to realize the weight distribution of different dimensions and positions of the expression area extracted by the network, focusing on the subtle difference feature information in the key points of the expression; the detail module is used to further extract the deep feature information. In order to obtain higher accuracy, a joint loss function is introduced to increase the out-of-class distance and reduce the intra-class distance to improve the accuracy of expression recognition. This paper applies this network to the datasets FER2013 and CK+. The experimental results show that the average recognition rate of this algorithm is respectively and 63.91%，97.98%the number of parameters is 11.34 M. Compared with VGG network, residual network, etc., this model not only improves the recognition rate but also reduces the redundant parameters.

Dual-grained Lightweight Strategy.

Removing redundant parameters and computations before the model training has attracted a great interest as it can effectively reduce the storage space of the model, speed up the training and inference of the model, and save energy consumption during the running of the model. In addition, the simplification of deep neural network models can enable high-performance network models to be deployed to resource-constrained edge devices, thus promoting the development of the intelligent world. However, current pruning at initialization methods exhibit poor performance at extreme sparsity. In order to improve the performance of the model under extreme sparsity, this paper proposes a dual-grained lightweight strategy-TEDEPR. This is the first time that TEDEPR has used tensor theory in the pruning at initialization method to optimize the structure of a sparse sub-network model and improve its performance. Specifically, firstly, at the coarse-grained level, we represent the weight matrix or weight tensor of the model as a low-rank tensor decomposition form and use multi-step chain operations to enhance the feature extraction capability of the base module to construct a low-rank compact network model. Secondly, unimportant weights are pruned at a fine-grained level based on the trainability of the weights in the low-rank model before the training of the model, resulting in the final compressed model. To evaluate the superiority of TEDEPR, we conducted extensive experiments on MNIST, UCF11, CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet datasets with LeNet, LSTM, VGGNet, ResNet and Transformer architectures, and compared with state-of-the-art methods. The experimental results show that TEDEPR has higher accuracy, faster training and inference, and less storage space than other pruning at initialization methods under extreme sparsity.

Pruning convolutional neural networks for inductive conformal prediction

Neural network pruning is a popular approach to reducing model storage size and inference time by removing redundant parameters in the neural network. However, the uncertainty of predictions from pruned models is unexplored. In this paper we study neural network pruning in the context of conformal predictors (CP). The conformal prediction framework built on top of machine learning algorithms supplements their predictions with reliable uncertainty measure in the form of prediction sets, under the independent and identically distributed assumption on the data. Convolutional neural networks (CNNs) have complicated architectures and are widely used in various applications nowadays. Therefore, we focus on pruning CNNs and, in particular, filter-level pruning. We first propose a brute force method that estimates the contribution of a filter to the CP’s predictive efficiency and removes those with the least contribution. Given the computation inefficiency of the brute force method, we also propose the Taylor expansion to approximate the filter’s contribution. Furthermore, we improve the global pruning method by protecting the most important filters within each layer from being pruned. In addition, we explore the ConfTr loss function which is optimized to yield maximal CP efficiency in the context of neural network pruning. We have conducted extensive experimental studies and compared the results regarding the trade-offs between predictive efficiency, computational efficiency, and network sparsity. These results are instructive for deploying pruned neural networks with applications using conformal prediction where reliable predictions and reduced computational cost are relevant, such as in safety-critical applications.

An unconstrained and non-redundant identification method of geometric errors and compensation of machine tools by X-AX Laserbar

Efficient and accurate measurement and identification of geometric errors are crucial for improving the precision of CNC machine tools. The X-AX Laserbar, as a novel tool for indirect measurement, has not been extensively studied for the identification of geometric errors in machine tools. In this paper, the geometric error model for a three-axis machine tool is established to illustrate the multilateration measurement principle of the laserbar, and a non-redundant and unconstrained identification method is proposed to identify these geometric errors. This method avoids the use of redundant parameters and additional constraints by employing pose error twists to describe the geometric errors. These pose error twists are identified in a transitional coordinate system, and then the geometric errors will be identified in the machine coordinate system by deriving the relationship between the pose errors and geometric errors. The proposed method is validated with the VMC-850E three-axis machine tool. The geometric error measurement using a laserbar is completed in about 40 min, showing great efficiency. The experimental results indicate that the proposed method is capable of accurately identifying the 17 geometric errors required for error compensation. The identified geometric errors are then applied to the machine tool’s accuracy improvement through error compensation. The results show that the actual geometric errors are controlled to a low level. The proposed method can efficiently measure the geometric errors of three-axis machine tools and contribute significantly to improving their geometric accuracy.

Learning Constrained Dynamic Correlations in Spatiotemporal Graphs for Motion Prediction.

Human motion prediction is challenging due to the complex spatiotemporal feature modeling. Among all methods, graph convolution networks (GCNs) are extensively utilized because of their superiority in explicit connection modeling. Within a GCN, the graph correlation adjacency matrix drives feature aggregation, and thus, is the key to extracting predictive motion features. State-of-the-art methods decompose the spatiotemporal correlation into spatial correlations for each frame and temporal correlations for each joint. Directly parameterizing these correlations introduces redundant parameters to represent common relations shared by all frames and all joints. Besides, the spatiotemporal graph adjacency matrix is the same for different motion samples, and thus, cannot reflect samplewise correspondence variances. To overcome these two bottlenecks, we propose dynamic spatiotemporal decompose GC (DSTD-GC), which only takes 28.6% parameters of the state-of-the-art GC. The key of DSTD-GC is constrained dynamic correlation modeling, which explicitly parameterizes the common static constraints as a spatial/temporal vanilla adjacency matrix shared by all frames/joints and dynamically extracts correspondence variances for each frame/joint with an adjustment modeling function. For each sample, the common constrained adjacency matrices are fixed to represent generic motion patterns, while the extracted variances complete the matrices with specific pattern adjustments. Meanwhile, we mathematically reformulate GCs on spatiotemporal graphs into a unified form and find that DSTD-GC relaxes certain constraints of other GC, which contributes to a better representation capability. Moreover, by combining DSTD-GC with prior knowledge like body connection and temporal context, we propose a powerful spatiotemporal GCN called DSTD-GCN. On the Human3.6M, Carnegie Mellon University (CMU) Mocap, and 3D Poses in the Wild (3DPW) datasets, DSTD-GCN outperforms state-of-the-art methods by 3.9%-8.7% in prediction accuracy with 55.0%-96.9% fewer parameters. Codes are available at https://github.com/Jaakk0F/DSTD-GCN.

A Multi-Model Architecture Based on Deep Learning for Longitudinal Available Overload Prediction of Commercial Subsonic Aircraft with Actuator Faults

Assessing the real-time longitudinal available overload onboard under fault conditions offers vital insights for the fault-tolerant reconfiguration and trajectory planning of commercial subsonic aircraft. After actuator failures in a commercial subsonic aircraft, its aerodynamic model undergoes changes. Traditional methods based on analytical models rely on precise aerodynamic models. However, due to the complexities of the flight environment and uncertainties in disturbances, establishing an accurate aerodynamic model after actuator failures is often challenging. Consequently, traditional methods can yield significant errors when evaluating the available overload under actuator faults. To address this, we introduce a multi-model architecture based on deep learning for the longitudinal available overload prediction of a commercial subsonic aircraft with actuator faults. For flight state data under different working conditions and different faults, Spearman correlation coefficient analysis and the gradient boosting decision tree (GBDT) algorithm are used to remove redundant feature parameters, thereby enhancing the training and prediction speed of the model while reducing the risk of overfitting. To meet prediction accuracy and speed demands, we employ the multi-layer perceptron (MLP) deep learning network to fully explore the environmental features, including uncertainties and disturbances, within the flight state, and the mapping relationships between the flight state and the available overload variations. We incorporate the light gradient boosting machine (LightGBM) and the categorical boosting (CatBoost) algorithms to enhance the model’s prediction speed and fuse it with a longitudinal available overload analytical model to elevate the model’s prediction accuracy, thereby achieving the real-time estimation of the commercial subsonic aircraft’s longitudinal available overload with actuator faults. The results demonstrate that the proposed method achieves a higher accuracy than traditional methods, with a relative error of less than 5%.

Fault detection method for transmission line components based on lightweight GMPPD-YOLO

Abstract This paper designs a lightweight high-precision transmission line component detection model, named grouped dense, monotonic self-regularized, and partial faster convolution, pruning, and distillation optimized—you only look once (GMPPD-YOLO), in transmission line inspection. It addresses the issue of low detection accuracy of target detection algorithms due to the complex background, large differences in target shape, location, texture, etc, as well as diversified and smaller defects in insulator and vibration hammer images taken by unmanned aerial vehicles from multiple angles. To enhance the model’s feature extraction capabilities in complex backgrounds and across different scales, the grouped dense C3 dense feature extraction module was designed, enabling the model to more effectively handle diverse defect forms. Simultaneously, the monotonic self-regularized pyramid pooling–fast (MSPPF) module is proposed to enhance the model’s capability to process multi-scale information. Additionally, the partial-faster C3 feature awareness module is designed to improve feature fusion performance, enhancing the model’s ability to perceive features at different scales. Finally, channel pruning was used to reduce redundant parameters, and knowledge distillation was employed to compensate for the accuracy loss caused by pruning. This approach further compressed the model size while ensuring its detection performance. The experimental results demonstrate that compared to the original YOLOv5s algorithm, the proposed GMPPD-YOLO algorithm achieves a reduction in parameters by 68.4%, a decrease in Giga floating-point operations per second by 58.2%, and a reduction in the model size by 66.4%, while achieving an increase in precision by 1%, mAP50 by 1.1%, and mAP95 by 0.4%. This confirms the significant potential of the GMPPD-YOLO algorithm for deployment in real-time drone-based power transmission line inspections.

A feature extension and reconstruction method with incremental learning capabilities under limited samples for intelligent diagnosis

Existing data-driven Intelligent Fault Diagnosis (IFD) methods often struggle with continuous learning and integrating new diagnostic knowledge. These methods necessitate a complete retraining process when new fault data emerges. Furthermore, collecting sufficient data for model training is impractical in real-world applications. This study proposes a Feature Extension and Reconstruction (FER) method for model incremental updating with limited samples. FER employs scalable and modular feature extractors for continuous representation learning of new fault modes. In addition, a domain mapping algorithm based on optimal transport is devised to leverage old classification nodes in aiding new nodes’ training, enhancing models’ generalization performance under limited samples. A feature reconstruction technique is also developed to compress the expanded multi-backbone model, eliminating redundant parameters and dimensions for adaptability in subsequent incremental tasks. Validated on simulated transmission systems and real planetary gearboxes, FER demonstrates a better stability-plasticity trade-off in incremental tasks with limited training samples, offering a novel solution for adaptive IFD model updates in sample-constrained scenarios.

Self-distillation enhanced adaptive pruning of convolutional neural networks

Convolutional neural networks (CNNs) suffer from issues of large parameter size and high computational complexity. To address this, we propose an adaptive pruning algorithm based on self-distillation. The algorithm introduces a trainable parameter for each channel to control channel pruning and integrates the pruning process into network training, enabling pruning and fine-tuning in a single training iteration to derive the final pruned model. Moreover, this framework requires only a single overall pruning rate to achieve adaptive pruning for each layer, avoiding tedious hyperparameter tuning for a less iterative, simple and efficient pruning process. Additionally, self-distillation is utilized in the pruning algorithm, leveraging the knowledge from the pretrained CNN to guide its own pruning process, facilitating the recovery from performance degradation caused by pruning and achieving higher accuracy. Extensive pruning experiments on various CNN models over different datasets demonstrate that at least 75% of redundant parameters can be reduced without sacrificing model accuracy.

PCB plug-in solder joint defect detection method based on coordinated attention-guided information fusion

Printed Circuit Boards (PCBs) are the foundational component of electronic devices, and the detection of PCB defects is essential for ensuring the quality control of electronic products. Aiming at the problem that the existing PCB plug-in solder defect detection algorithms cannot meet the requirements of high precision, low false alarm rate, and high speed at the same time, this paper proposes a method based on spatial convolution pooling and information fusion. Firstly, on the basis of YOLOv3, an attention-guided pyramid structure is used to fuse context information, and multiple convolutions of different size are used to explore richer high-level semantic information; Secondly, a coordinated attention network structure is introduced to calibrate the fused pyramidal feature information, highlighting the important feature channels, and reducing the adverse impact of redundant parameters generated by feature fusion; Finally, the ASPP (Atrous Spatial Pyramid Pooling) structure is implemented in the original Darknet53 backbone feature extraction network to acquire multi-scale feature information of the detection targets. With these improvements, the average detection accuracy of the enhanced network has been elevated from 94.45 to 96.43%. This experiments shows that the improved network is more suitable for PCB plug-in solder defect detection applications.

A Lightweight Insulator Defect Detection Model Based on Drone Images

With the continuous development and construction of new power systems, using drones to inspect the condition of transmission line insulators has become an inevitable trend. To facilitate the deployment of drone hardware equipment, this paper proposes IDD-YOLO (Insulator Defect Detection-YOLO), a lightweight insulator defect detection model. Initially, the backbone network of IDD-YOLO employs GhostNet for feature extraction. However, due to the limited feature extraction capability of GhostNet, we designed a lightweight attention mechanism called LCSA (Lightweight Channel-Spatial Attention), which is combined with GhostNet to capture features more comprehensively. Secondly, the neck network of IDD-YOLO utilizes PANet for feature transformation and introduces GSConv and C3Ghost convolution modules to reduce redundant parameters and lighten the network. The head network employs the YOLO detection head, incorporating the EIOU loss function and Mish activation function to optimize the speed and accuracy of insulator defect detection. Finally, the model is optimized using TensorRT and deployed on the NVIDIA Jetson TX2 NX mobile platform to test the actual inference speed of the model. The experimental results demonstrate that the model exhibits outstanding performance on both the proprietary ID-2024 insulator defect dataset and the public SFID insulator dataset. After optimization with TensorRT, the actual inference speed of the IDD-YOLO model reached 20.83 frames per second (FPS), meeting the demands for accurate and real-time inspection of insulator defects by drones.

Detection Method of Stator Coating Quality of Flat Wire Motor Based on Improved YOLOv8s.

The stator of a flat wire motor is the core component of new energy vehicles. However, detecting quality defects in the coating process in real-time is a challenge. Moreover, the number of defects is large, and the pixels of a single defect are very few, which make it difficult to distinguish the defect features and make accurate detection more difficult. To solve this problem, this article proposes the YOLOv8s-DFJA network. The network is based on YOLOv8s, which uses DSFI-HEAD to replace the original detection head, realizing task alignment. It enhances joint features between the classification task and localization task and improves the ability of network detection. The LEFG module replaces the C2f module in the backbone of the YOLOv8s network that reduces the redundant parameters brought by the traditional BottleNeck structure. It also enhances the feature extraction and gradient flow ability to achieve the lightweight of the network. For this research, we produced our own dataset of stator coating quality regarding flat wire motors. Data augmentation technology (Gaussian noise, adjusting brightness, etc.) enriches the dataset, to a certain extent, which improves the robustness and generalization ability of YOLOv8s-DFJA. The experimental results show that in the performance of YOLOv8s-DFJA compared with YOLOv8s, the mAP@.5 index increased by 6.4%, the precision index increased by 1.1%, the recall index increased by 8.1%, the FPS index increased by 9.8FPS/s, and the parameters decreased by 3 Mb. Therefore, YOLOv8s-DFJA can be better realize the fast and accurate detection of the stator coating quality of flat wire motors.

TAKAGI–SUGENO–KANG FUZZY SYSTEM MODELING BASED ON LOW-RANK SPARSE SUBSPACE LEARNING FOR MOTOR IMAGERY ELECTROENCEPHALOGRAM SIGNAL CLASSIFICATION

The classification of electroencephalogram (EEG) signals derived from motor imagery (MI) has always been a hot topic in the field of brain–computer interfaces. Due to its ability to handle the nonstationary and uncertain information contained in EEG signals, the Takagi–Sugeno–Kang fuzzy system (TSK-FS) has become an advantageous classification algorithm. To train a fuzzy system with strong discrimination capabilities from EEG data interspersed with redundant information, this paper proposes a TSK-FS modeling method based on low-rank sparse subspace learning (TSK-LSSL). This method focuses on consequent parameter learning, which transforms the traditional consequent parameter learning strategy into low-rank subspace and sparse subspace learning processes. Low-rank subspace learning is used to mine the global structural information of data and effectively reduce the number of fuzzy rules. During sparse subspace learning, [Formula: see text]-norm regularization is used to constrain the consequent parameters and causes the number of redundant consequent parameters to be zero, thereby simplifying the fuzzy rules. In addition, a local boundary term based on graph matrices is embedded into the objective function to mine the local structural information of the given data. TSK-LSSL simplifies the number of rules and the consequent part of the fuzzy rules. It exhibits good classification performance on two BCI Competition databases.

Intelligent identification method for pipeline ultrasonic internal inspection

ABSTRACT In pipeline ultrasonic testing, the accurate identification of defect types is a prerequisite for quantifying defect sizes. To this end, this paper proposes a defect identification algorithm based on the combination of Wavelet Packet Transform (WPT), Distance Evaluation Technique (DET) and GA-SVM to precisely identify defect boundaries. First, the ultrasonic echo signals are processed using WPT to more accurately extract defect information. Then, DET is employed to eliminate redundant feature parameters, which are sorted by sensitivity to determine the input feature parameters for the GA-SVM. On this basis, a self-developed ultrasonic testing experimental system is used to collect ultrasonic echo signals from Q235 steel pipes with cylindrical, spherical, and conical defects at different spatial orientations of the probe and the pipe. After processing and identifying the signals, it was found that the proposed algorithm achieves an average identification accuracy of 99% for different types of defects, thereby realising automatic identification of pipeline defects. The results provide new ideas and methods for the intelligent identification of volumetric defect types in pipeline ultrasonic internal testing and promote the development of industrial-grade ultrasonic internal detectors.

A high-speed YOLO detection model for steel surface defects with the channel residual convolution and fusion-distribution

Abstract Accurately and efficiently detecting steel surface defects is a critical step in steel manufacturing. However, the compromise between the detection speed and accuracy remains a major challenge, especially for steel surface defects with large variations in the scale. To address the issue, an improved you only look once (YOLO) based detection model is proposed through the reinforcement of its backbone and neck. Firstly, for the reduction of the redundant parameters and also the improvement of the characterization ability of the model, an effective channel residual structure is adopted to construct a channel residual convolution module and channel residual cross stage partial module as components of the backbone network, respectively. They realize the extraction of both the shallow feature and multi-scale feature simultaneously under a small number of convolutional parameters. Secondly, in the neck of YOLO, a fusion-distribution strategy is employed, which extracts and fuses multi-scale feature maps from the backbone network to provide global information, and then distributes global information into local features of different branches through an inject attention mechanism, thus enhancing the feature gap between different branches. Then, a model called CRFD-YOLO is derived for the steel surface defect detection and localization for the situations where both speed and accuracy are demanding. Finally, extensive experimental validations are conducted to evaluate the performance of CRFD-YOLO. The validation results indicate that CRFD-YOLO achieves a satisfactory detection performance with a mean average precision of 81.3% on the NEU-DET and 71.1% on the GC10-DET. Additionally, CRFD-YOLO achieves a speed of 161 frames per second, giving a great potential in real-time detection and localization tasks.

Deep incomplete multi-view clustering via attention-based direct contrastive learning

Incomplete Multi-View Clustering (IMVC), enhanced by contrastive learning, stands out in unsupervised learning for its notable performance. However, it faces challenges: over-reliance on additional projection heads to avoid dimensionality collapse, leading to redundant parameters, and the risk of encoder-derived features, which merge view-specific information, misleading the learning of common semantics due to simultaneous learning and reconstruction in the same feature space. To address these issues, we propose a novel framework for incomplete multi-view contrastive clustering. This framework employs an encoder network with a self-attention mechanism, allowing both reconstruction loss and contrastive loss to act on the learned feature vectors and their sub-vectors, respectively. This approach effectively mitigates the impact of extraneous private information. By leveraging sub-vectors for consistency learning, our model directly refines the latent feature subspace, thus circumventing dimensionality collapse without the dependence on projection heads. Additionally, our method incorporates a cross-view prediction mechanism to recuperate missing information in incomplete datasets. Comprehensive experiments on public datasets demonstrate that our method achieves state-of-the-art clustering performance.

TinyCheXReport: Compressed deep neural network for Chest X-ray report generation

Increase in chest x-ray (CXR) imaging tests has burdened radiologists, thereby posing significant challenges in writing radiological reports on time. Although several deep learning-based automatic report generation methods have been developed till date, most are over-parameterized. For deployment on edge devices with constrained processing power or limited resources, over-parameterized models are often too large. This paper presents a compressed deep learning-based model that is 30% space efficient as compared to the non-compressed base model while both have comparable performance. The model comprising of VGG19 and hierarchical long short-term memory (LSTM) equipped with contextual word embedding layer is used as the base model. The redundant weight parameters are removed from the base model using unstructured one-shot pruning. To overcome the performance degradation, the lightweight pruned model is fine-tuned over publically available OpenI dataset. The quantitative evaluation metric scores demonstrate that proposed model surpasses the performance of state-of-the-art models. Additionally, the proposed model, being 30% space efficient, is easily deployable in resource-limited settings. Thus, this study serves as baseline for development of compressed models to generate radiologicalreports from CXR images.

Transformer Discharge Carbon-Trace Detection Based on Improved MSRCR Image-Enhancement Algorithm and YOLOv8 Model.

It is difficult to visually detect internal defects in a large transformer with a metal closure. For convenient internal inspection, a micro-robot was adopted, and an inspection method based on an image-enhancement algorithm and an improved deep-learning network was proposed in this paper. Considering the dim environment inside the transformer and the problems of irregular imaging distance and fluctuating supplementary light conditions during image acquisition with the internal-inspection robot, an improved MSRCR algorithm for image enhancement was proposed. It could analyze the local contrast of the image and enhance the details on multiple scales. At the same time, a white-balance algorithm was introduced to enhance the contrast and brightness and solve the problems of overexposure and color distortion. To improve the target recognition performance of complex carbon-trace defects, the SimAM mechanism was incorporated into the Backbone network of the YOLOv8 model to enhance the extraction of carbon-trace features. Meanwhile, the DyHead dynamic detection Head framework was constructed at the output of the YOLOv8 model to improve the perception of local carbon traces with different sizes. To improve the defect target recognition speed of the transformer-inspection robot, a pruning operation was carried out on the YOLOv8 model to remove redundant parameters, realize model lightness, and improve detection efficiency. To verify the effectiveness of the improved algorithm, the detection model was trained and validated with the carbon-trace dataset. The results showed that the MSH-YOLOv8 algorithm achieved an accuracy of 91.80%, which was 3.4 percentage points higher compared to the original YOLOv8 algorithm, and had a significant advantage over other mainstream target-detection algorithms. Meanwhile, the FPS of the proposed algorithm was up to 99.2, indicating that the model computation and model complexity were successfully reduced, which meets the requirements for engineering applications of the transformer internal-inspection robot.