Abstract:Given the challenges posed by photovoltaic component damage detection and the high demands placed on human and computational resources by existing detection technologies, an improved lightweight model named YOLOv8-DM was proposed based on YOLOv8n. The integration of Electroluminescence imaging with object detection methods was implemented to achieve PV defect detection. Innovative components were introduced, including a Dynamic Scale Feature Pyramid Network and an Inverted Residual Multiscale Attention Mechanism, along with a Ghost module enhanced by dynamic convolution. These modifications were specifically designed to address the deficiencies observed in the YOLOv8n model regarding feature representation and multiscale object recognition, which enhanced fine-grained detection capabilities and reduced computational complexity. When evaluated on the augmented PVEL-AD dataset, the model demonstrated an improvement of 3% in recall rate and 3.3% in mAP50 compared to the baseline model, with a 34% reduction in parameter count and a 20% decrease in computational demand. The optimized architecture was validated to effectively meet the practical requirements for high-accuracy PV defect detection with lower computational costs.

Abstract:Long-sequence autonomous manipulation capability becomes one of the bottlenecks hindering the practical application of intelligent robots. To address the diverse long-sequence operation skill requirements faced by robots in complex scenarios, an efficient and robust asymmetric Actor-Critic reinforcement learning method was proposed. This approach aims to solve the challenges of high learning difficulty and complex reward function design in long-sequence tasks. By integrating multiple Critic networks to collaboratively train a single Actor network, and introducing GAIL (generative adversarial imitation learning) to generate intrinsic rewards for the Critic network, the learning difficulty of long-sequence tasks was reduced. On this basis, a two-stage learning method was designed, utilizing imitation learning to provide high-quality pre-trained behavior policies for reinforcement learning, which not only improves learning efficiency but also enhances the generalization performance of the policy. Simulation results for long-sequence autonomous task execution in a chemical laboratory demonstrate that the proposed method significantly improves the learning efficiency of robot long-sequence skills and the robustness of behavior policies.

Abstract:Source region integration algorithm is a kind of data processing method to extract noise characteristics of each component in wind tunnel test model. To solve the interference problem of integration results caused by sound source sidelobe outside the target integration, a Partial-CLEAN integration algorithm is proposed. This algorithm divides the model region into target and non-target regions, and uses the CLEAN algorithm to find the strongest sound source positions at each frequency point in the non-target region. It iteratively removes interference from non-target region sound sources on the target region, thereby achieving more accurate integration results. Through simulation and wind tunnel test data analysis, the Partial-CLEAN integration algorithm can effectively separate mutually interfering noise sources, especially sound sources of 3kHz and below, extract more accurate target sound sources, and provide a new tool for aerodynamic noise analysis.

Abstract:To address the problem of how to faithfully map neural networks to resource-constrained embedded devices, a mixed-precision quantization method for convolutional neural networks based on layer sensitivity analysis is proposed. The sensitivity of convolutional layer parameters is measured by calculating the average trace of the Hessian matrix, providing a basis for bit-width allocation. A layer-wise ascending-descending approach is employed for bit-width allocation, ultimately achieving mixed-precision quantization of the network model.Experimental results demonstrate that compared to the fixed-precision quantization methods DoReFa and LSQ+, the proposed mixed-precision quantization method improves recognition accuracy by 10.2% and 1.7%, respectively, at an average bit-width of 3 bits. When compared to other mixed-precision quantization methods, the proposed approach achieves over 1% higher recognition accuracy. Additionally, noise-injected training effectively enhances the robustness of the mixed-precision quantization method, improving recognition accuracy by 16% under a noise standard deviation of 0.5.

Abstract:non-coplanar flyby is an effective method for monitoring important satellites in geosynchronous orbits. Taking the scenario of patrollong the geosynchronous belt with an elliptical orbit as an example,the paper proposed a method for solving the minimal orbital intersection distance(MOID) between any two elliptic orbits,proving that the optimal flyby point for the patrollong satellite is the Asending Node.The paper used ATK to model the mission scenario,investigating the variations in the optimal flyby point position in J2 gravitational model.The paper studied the variations in terminal constraints such as relative distance and sun phase angle at flyby point in the situation of different local times and different flyby directions.The paper proposed a transfer strategy that satisfies the maximum transfer time and terminal constraints,using the Astrogator in ATK,the mission parameters for the mission under different gravitational model assumptions are solved.The paper analyzed mission parameter corresponding to different true anomaly angles in two-body model,validating the effectiveness of the strategy and providing a basis for planning multiple patrol sequences.

Abstract:Due to the limitations of the number of satellites, scale and manoeuvrability, it is difficult for the existing earth observation constellations to respond quickly to emergency needs of a high degree of randomness. In order to meet the demand for rapid design of emergency earth observation constellations, an emergency earth observation constellation design method was proposed. Based on the secondary development and coverage analysis function of the Aerospace Tool Kit, the method adopted a one-dimensional data hierarchical clustering method to group ground targets, and then applied the differential evolution algorithm to optimize the restricted walker subconstellation configuration for the target groups, and finally generated the restricted hybrid walker earth observation constellation. Simulation results demonstrate that the method can rapidly generate emergency constellations compared with conventional walker constellations and violent optimisation results, and minimize satellite deployment quantity while ensuring the effective completion of the earth observation mission.

Abstract:Gear faults manifest in the frequency spectrum as multi-order modulation sideband clusters centered on the meshing frequency and its higher-order harmonics, spaced by the gear rotation frequency. In order to automatically focus the fault side frequency components, a method of fault side frequency cluster extraction with penalty regression was proposed. The Adaptive sparse group lasso (ASGL) regression self-data-driven strategy was used to determine the penalty coefficient and update the spectrum weight online to find the fault sideband clusters. Based on the sideband weight coefficients obtained from Sparse Group Lasso (SGL) regression, a new index called the Sparse Group Lasso Sidebands Indicator (SSI) was proposed for the health monitoring of gear transmission systems, enabling the early fault warning and location of gear transmission systems. The results show that the proposed method can provide more accurate early fault detection and fault location results.

Abstract:To address the problem of inadequate self-supervised signal quality in contrastive learning models for sequential recommendation, a combinatorial enumeration and time-interval contrastive learning model was proposed. The model generated enhanced sequences which preserved temporal information through time-interval perturbation-based data augmentation. A combinatorial enumeration strategy was introduced to integrate user behavior and time-interval information, constructing multi-view augmented sequence pairs. The model employed a multi-head attention mechanism to encode user behavior sequences and optimized self-supervised signals through multi-task joint training, which improved overall performance. The proposed model is well-suited for scenarios with high data sparsity and uneven interaction behaviors, effectively addressing challenges in self-supervised signal modeling. Experimental results on three real-world datasets demonstrate that the model outperforms state-of-the-art contrastive learning models in terms of Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG). Specifically, HR@5 and NDCG@5 improve by 5.61% and 8.53%, respectively.

Abstract:Successful application of reinforcement learning in decision support, combinatorial optimization, and intelligent control has driven its exploration in complex industrial scenarios. However, existing reinforcement learning methods face challenges in adapting to graph-structured data in non-Euclidean spaces. Graph neural networks have demonstrated exceptional performance in learning graph-structured data. By integrating graphs with reinforcement learning, graph-structured data was introduced into reinforcement learning tasks, enriching knowledge representation in reinforcement learning and offering a novel paradigm for addressing complex industrial process problems. the research progress of graph reinforcement learning algorithms in industrial domains was systematically reviewed, summarized graph reinforcement learning algorithms from the perspective of algorithm architecture and extracted three mainstream paradigms, explored their applications in production scheduling, industrial knowledge graph reasoning, industrial internet, power system and other fields, and analyzed current challenges alongside future development trends in this field.

Abstract:In order to solve the problem that it is difficult to identify the forced motion intention of non-cooperative targets, An intention recognition method based on BiGRU network was proposed in this paper. The non-cooperative target was categorizes into five forced motion intentions: "Forced Round Fly-Around", "Forced Drip-drop Fly-Around", "Fixed-point Oscillating", "Line Approach" and "Hop Approach", and the forced motion intention maneuver information dataset of the non-cooperative target was established. Based on the maneuver time series information of the non-cooperative target after entering the observation range of our spacecraft, the BiGRU network was utilized to train on the potential correlation between the time series data and the forced motion intention, so as to realize the intention recognition of the non-cooperative target. The simulation results demonstrated that the detection accuracy of the BiGRU network-based forced motion intention recognition method for non-cooperative targets achieved 98.35%. This method can improve the ability to identify the intentions of non-cooperative targets and provide a technical reference for the safety of our spacecraft in orbit.