Huapeng Yu

Traditional maritime target element resolution, relying on manual experience and uniform sampling, lacks accuracy and efficiency in non-uniform sampling, missing data, and noisy scenarios. While large language models (LLMs) offer a solution, their general knowledge gaps with maritime needs limit direct application. This paper proposes a fine-tuned LLM-based adaptive optimization method for non-uniform sampling maritime target element resolution, with three key novelties: first, selecting Doubao-Seed-1.6 as the base model and conducting targeted preprocessing on maritime multi-source data to address domain adaptation gaps; second, innovating a “Prefix tuning + LoRA” hybrid strategy (encoding maritime rules via Prefix tuning, freezing 95% of base parameters via LoRA to reduce trainable parameters to <0.5%) to balance cost and performance; third, building a non-uniform sampling-model collaboration mechanism, where the fine-tuned model dynamically adjusts the sampling density via semantic understanding to solve random sampling’s “structural information imbalance”. Experiments in close, away, and avoid scenarios (vs. five control models including original LLMs, rule-only/models, and ChatGPT-4.0) show that the proposed method achieves a comprehensive final score of 0.8133—37.1% higher than the sub-optimal data-only model (0.5933) and 87.7% higher than the original general model (0.4333). In high-risk avoid scenarios, its Top-1 Accuracy (0.7333) is 46.7% higher than the sub-optimal control, and Scene-Sensitive Recall (0.7333) is 2.2 times the original model; in close and away scenarios, its Top-1 Accuracy reaches 0.8667 and 0.9000, respectively. This method enhances resolution accuracy and adaptability, promoting LLM applications in navigation.

This paper aims to study the design of deep learning adaptive algorithms for different types of inertial navigation systems. Underwater inertial navigation plays a crucial role in fields such as ocean exploration, but improving its accuracy faces challenges, and combined navigation methods are often adopted. Deep learning is widely applied in inertial navigation, which can be used for sensor data processing, positioning, attitude estimation, and can also be combined with other navigation technologies. Based on previous research, this paper delves into the characteristics of different inertial navigation devices, combines with large language models, and proposes an LLM-based adaptive parameter selection expert system. This system can select appropriate deep learning network parameters based on parameters such as the material, accuracy, and service life of the inertial navigation device to ensure the matching between the model and the inertial navigation device. The network design is divided into a displacement prediction module and a heading angle prediction module, and different targeted designs are adopted. The experiment was conducted in the Yellow Sea, equipped with strapdown fiber optic inertial navigation systems and strapdown laser inertial navigation systems, and targeted training was carried out according to the characteristics of the two sets of systems. The fiber optic inertial navigation system has an earlier factory time and a large data scale; the laser inertial navigation system has a recent factory time and high accuracy. The experimental results show that the error is significantly suppressed after adopting the deep learning algorithm, and it is feasible to use the expert system to drive the setting of targeted parameters, which has the value of promotion and research.

Constrained by the significant attenuation of electromagnetic waves in water, inertial navigation has become the core autonomous and controllable navigation method for underwater navigation. However, traditional inertial navigation methods suffer from increasing errors in inertial devices over time, leading to rapid divergence and loss of navigational capability in the absence of external correction source information input. Moreover, high-precision inertial navigation systems are often expensive and difficult to apply in consumer-grade Autonomous Underwater Vehicles (AUVs). This paper conducts an in-depth study on the time factor that has been overlooked in traditional inertial navigation, focusing on the inherent errors of inertial devices, vehicle attitude changes during navigation, and error propagation in positioning. The study explores the application of the time factor in deep learning methods for inertial navigation, proposing a pure inertial deep learning navigation method based on time constraints. This method incorporates time factors neglected in traditional navigation methods and designs multiple motion-state self-switching models tailored to different vehicle motion states. Finally, the effectiveness of the method is validated through actual long-range sea trials, with experimental results demonstrating higher prediction accuracy compared to other mainstream methods and effective suppression of error divergence.

Inertial navigation systems (INSs) are a topical solution in underwater navigation. Although appealing due to their ability to estimate pose without external information, INS suffer from compounding position errors due to bias and random noise. In general, INSs require the assistance of other positioning devices to achieve satisfactory positioning results. To solve these problems, this paper proposes an ego-motion estimation framework with an inertial measurement unit (IMUs) and magnetic compass based on the deep learning theory. The main idea is to estimate the displacement of vehicles from the IMU data in the time window and combine this with magnetic compass headings to reconstruct the trajectories of the vehicles. The pre-integration technology is used to process raw IMU data, which mathematically separates the dependence of traditional inertial algorithms based on the initial value. Then convolutional neural networks (CNN) and attention hybrid networks are used to estimate the displacement of vehicles. In addition, the framework leverages the backpropagation neural network to fuse the magnetic heading and IMU measurements to obtain an accurate heading. Compared with other deep learning methods, the proposed method reduces computational complexity and improves position accuracy. Eventually, the accuracy of the proposed method is verified in the sea trail. The results show that the maximum value of absolute trajectory errors accounts for 12.8% of the distance in severe sea conditions and 6.38% in usual sea conditions.

The pure inertial navigation system, crucial for autonomous navigation in GPS-denied environments, faces challenges of error accumulation over time, impacting its effectiveness for prolonged missions. Traditional methods to enhance accuracy have focused on improving instrumentation and algorithms but face limitations due to complexity and costs. This study introduces a novel device-level redundant inertial navigation framework using high-precision accelerometers combined with a neural network-based method to refine navigation accuracy. Experimental validation confirms that this integration significantly boosts navigational precision, outperforming conventional system-level redundancy approaches. The proposed method utilizes the advanced capabilities of high-precision accelerometers and deep learning to achieve superior predictive accuracy and error reduction. This research paves the way for the future integration of cutting-edge technologies like high-precision optomechanical and atom interferometer accelerometers, offering new directions for advanced inertial navigation systems and enhancing their application scope in challenging environments.

Underwater acoustic target recognition(UATR)is a challenging task Due to the high cost of sampling data, it is difficult to build a large-scale dataset.In this study,a new method based on transfer learning with a VGG16 model(Transfer-VGG16) is developed in which three-dimensional(3-D)data is used as the input.Use the dataset obtained in real scenarios to validate the proposed method.Experiments have shown that Transfer-VGG16 model works well when the observation data is scarce.

Forest fire has the characteristics of sudden and destructive, which threatens safety of people’s life and property. Automatic detection and early warning of forest fire in the early stage is very important for protecting forest resources and reducing disaster losses. Unmanned forest fire monitoring is one popular way of forest fire automatic detection. However, the actual forest environment is complex and diverse, and the vision image is affected by various factors easily such as geographical location, seasons, cloudy weather, day and night, etc. In this paper, we propose a novel fire detection method called Fire-PPYOLOE. We design a new backbone and neck structure leveraging large kernel convolution to capture a large arrange area of reception field based on the existing fast and accurate object detection model PP-YOLOE. In addition, our model maintains the high-speed performance of the single-stage detection model and reduces model parameters by using CSPNet significantly. Extensive experiments are conducted to show the effectiveness of Fire-PPYOLOE from the views of detection accuracy and speed. The results show that our Fire-PPYOLOE is able to detect the smoke- and flame-like objects because it can learn features around the object to be detected. It can provide real-time forest fire prevention and early detection.

The geomagnetic field constitutes a massive fingerprint database, and its unique structure provides potential position correction information. In recent years, particle filter technology has received more attention in the context of robot navigation. However, particle degradation and impoverishment have constrained navigation systems’ performance. This paper transforms particle filtering into a particle-matching positioning problem and proposes a geomagnetic localization method based on an adaptive fission particle filter. This method employs particle-filtering technology to construct a geomagnetic matching positioning model. Through adaptive particle fission and sampling, the problem of particle degradation and impoverishment in traditional particle filtering is solved, resulting in improved geomagnetic matching positioning accuracy. Finally, the proposed method was tested in a sea trial, and the results show that the proposed method has a lower positioning error than traditional particle-filtering and intelligent particle-filtering algorithms. Under geomagnetic map conditions, an average positioning accuracy of about 546.44 m is achieved.

Autonomous navigation technology is the key technology for Autonomous Underwater Vehicle (AUV) to achieve automated, intelligent operation and task processing. Inertial navigation technology is the core of autonomous navigation technology for AUV. Traditional inertial navigation technology has been developed for many years, and it is necessary to find new breakthroughs. Deep learning can automatically select and extract key features of input data, which has been widely used in image recognition, speech recognition, natural language processing and other fields, and has good results in processing sequential data such as text and speech. Inertial navigation data clearly belongs to this type of data, and many scholars in the industry have conducted related research and design, and found that deep neural network models can be used to calibrate the noise of inertial sensors, reduce the drift of inertial navigation mechanisms, and fuse inertial information with other sensor information, with good effects in solving the prediction and error suppression of inertial navigation during long-term underwater voyages. This article provides a comprehensive review of deep learning-based inertial navigation for AUV, including the latest research progress and development trend direction.

As a new type of ultra-high precision inertial device, the optomechanical accelerometer can change the current situation of insufficient composition requirements, accuracy, volume and cost adaptation of the accelerometer relative to the inertial system to a certain extent. However, the research on the optomechanical accelerometer is still in the initial stage. Variations in the optical parameters, structural parameters and optical trap environment of the optomechanical accelerometer system can affect its measurement performance. In this paper, the influence of the laser power, the microsphere radius, the centroid temperature, and the pressure on the sensing performance of the system is discussed. Through simulation results, the theoretical accuracy, measuring range, bandwidth and other key parameters are analyzed, which can provide indicators for the prototype development, the manufacturing process, and the data processing.