Bio-inspired optimisation methods have been widely applied to complex real-world problems, particularly in low-carbon power and energy systems, where optimisation tasks often involve high-dimensional, constrained and mixed-integer characteristics. Traditional approaches struggle with these challenges due to nonconvexity, nonlinearity and computational complexity. This paper provides a comprehensive review of bio-inspired optimisation techniques applied to key low-carbon energy problems, including economic load dispatch, unit commitment, optimal power flow, distributed generation planning, heat exchanger design, and parameter estimation for PEM fuel cells and solar cell models. By analysing the strengths and limitations of existing methods, we highlight their effectiveness in addressing computational efficiency, constraint handling and convergence behaviour. The paper also identifies research gaps and discusses future directions, providing a structured reference for algorithm developers and practitioners. This review aims to enhance the adoption and refinement of bio-inspired optimisation techniques for sustainable energy solutions.

This paper presents HealthNet, a novel framework for the dynamic optimisation of healthcare transportation networks using multi-agent reinforcement learning. HealthNet leverages a spatiotemporal dependency module to capture complex spatiotemporal relationships in healthcare demand and resource allocation patterns, combined with centralised training and a decentralised execution approach. The system is modelled as a Markov game and solved using a deep reinforcement learning algorithm. Extensive simulations demonstrate that HealthNet outperforms eight state-of-the-art baseline methods across multiple network configurations and evaluation metrics. In a 4 × 4 grid network, HealthNet reduces average waiting times by 47.6% compared to model predictive control and 22.1% compared to the best-performing baseline. Traffic congestion rates are reduced to 16.7% compared to 42.3% for the worst baseline and 23.1% for the best baseline. Under irregular network topologies with stochastic disruptions, including demand surges and vehicle unavailability, HealthNet maintains superior performance with 42.1% lower average waiting time and 51.1% improvement in peak response times compared to competing approaches. These findings indicate that HealthNet can enhance both efficiency and resilience in healthcare transportation systems, potentially improving patient outcomes in complex urban environments.

Intrinsic motivation serves as the predominant paradigm of exploration in reinforcement learning. In pursuit of an informative and robust state representation, the behavioural metric groups behaviourally equivalent states together, which share the same single-step reward and transition distribution. However, due to the presence of uninformative rewards and the dynamic nature of procedurally generated environments, these behavioural metric-based approaches could limit the effectiveness of the learnt state representations, potentially leading to a representation collapse and an ineffective exploration. Therefore, a more comprehensive and generalisable behavioural metric is needed to overcome the above issues. In this work, we approach the exploration problem from a novel perspective, extending beyond the conventional single-step assessments to encompass a long-term consideration of the whole trajectory. Specifically, we propose a novel trajectory-level behavioural metric (TBM) that exploits temporal dependencies of the trajectory and captures the underlying sequential information of behaviour patterns. To achieve an effective trajectory representation for exploration, we develop apivotal state identifier (PSI) and a trajectory return estimator (TRE) to distinguish the diverse contributions of individual states in the trajectory. Moreover, an auxiliary representation regulariser is developed to promote the diversity and informativeness of the trajectory representation, mitigating the risk of representation mode collapse. Extensive experiments and empirical analysis conducted on procedurally generated environments showcase the superior performance of our proposed framework.

Adversarial jamming attacks have increased on communication systems, causing distortion and threatening transmissions. Typical attacks rely on traditional, well-defined cryptographic protocols and frequency-hopping techniques. Nevertheless, these techniques become vulnerable when facing intelligent jammers. To address this issue, we introduce a new framework that integrates Siamese neural networks with a dual-probability-attention mechanism (DPAM) to provide reliable anti-jamming communication and robust protection. This framework contains several components, which are (1) twin neural networks to execute coordinated cryptographic adaptation operation using a contrastive learning approach, (2) a DPAM module to analyse signals using probability encoding and dual temporal-spectral attention to enhance accurate recognition, (3) adversarial training to counter growing attack patterns and (4) a lightweight neural encryption module that is developed to provide real-time operation. Internal DPAM architecture combines probability distributions with Bayesian attention fusion. This combination increases the detection by 23% when compared to other attention mechanisms. Conducted simulation evaluations on a public dataset shows that the frameworks reached an accuracy of 98.7%, whereas other reinforcement learning (RL) methods achieved 82%. In addition, 45% reduction in latency was reached when compared to frequency-hopping solutions. Furthermore, the solution got up to 96% resilience against attacks.

With the advancement of brain–computer interfaces (BCI), motor imagery (MI) electroencephalogram (EEG) decoding can greatly benefit from spatial filtering features derived from common spatial patterns (CSP). However, CSP-based features often exhibit high redundancy and intersubject variability. These limitations make the feature selection methods based on sparse learning difficult to effectively balance the heterogeneous contributions of different temporal and spatial components. Moreover, these models tend to prioritise features with larger coefficients, potentially overlooking intrinsic feature importance and compromising the quality of the selected feature subset. To address these issues, we propose an Adaptive Sparse Group Lasso (ASGL) method for structured feature selection, designed to enhance discriminative CSP features whilst suppressing irrelevant components. The proposed method partitions EEG signals into consecutive segments using a sliding window, treating each as a separate feature group. Benefiting from this, the importance of features at both the group level and the within-group level can be effectively quantified through mutual information and copula mutual information, thereby assigning adaptive weights for selective penalisation within the model. This weight construction strategy preserves important features from relevant time intervals and frequency bands. The resulting optimization problem is solved efficiently via the alternating direction method of multipliers (ADMM). Evaluations on simulated and real-world datasets demonstrate that the proposed ASGL outperforms existing methods.

To enable autonomous operations in complex industrial environments, this paper proposes retrieval-augmented generation with large language models for robotic operations (RAGLRO), a robotic framework specifically designed for power switchgear operation tasks. The system integrates multimodal perception with high-level semantic reasoning and task-level action generation. A depth camera captures the environmental context, which is processed by visual modules to perform object detection and pose detection. The perception outputs are formulated into structured prompts and provided to a large language model (LLM) equipped with a retrieval-augmented generation (RAG) mechanism. The RAG component enables the LLM to dynamically access a task-specific knowledge base, including operation manuals, safety protocols and historical mission data, thereby enhancing contextual understanding and reasoning precision. Based on the retrieved knowledge and current environmental perception, the LLM selects and sequences callable action functions from a predefined robotic action library to generate executable robot control commands. A dedicated dataset for power switchgear operations is also constructed to support robust visual perception, containing annotated images for object detection and pose detection tasks. Experimental results demonstrate that RAGLRO achieves high task success rates and strong adaptability in real-world power maintenance scenarios, validating the effectiveness of integrating multimodal perception, LLM-based reasoning and RAG-grounded task planning within a unified robotic control framework.

Influence maximisation (IM) aims to select a small number of nodes that are able to maximise their influence in a network and covers a wide range of applications. Despite numerous attempts to provide effective solutions in simple networks, higher-order interactions between entities in various real-world systems are usually not taken into account. In this paper, we propose a versatile meta-heuristic approach, Hypergraph Genetic Algorithm (HGA), to tackle the IM problem in hypergraphs, which is based on the concept of genetic evolution. Systematic validations in synthetic and empirical hypergraphs under both simple and complex hypergraph-based contagion models indicate that HGA achieves universal and plausible performance compared to baseline methods. We explore the cause of the excellent performance of HGA through ablation studies and correlation analysis. The findings show that the solution of HGA is distinct from that of other prior methods. Moreover, a closer look at the local topological features of the seed nodes acquired by different algorithms reveals that the selection of seed nodes cannot be based on a single topological characteristic but should involve a combination of multiple topological features to address the IM problem.

Efficient exploration is critical in handling sparse rewards and partial observability in deep reinforcement learning. However, most existing intrinsic reward methods based on novelty rely on single-step observations or Euclidean distances. These approaches struggle to capture trajectory-level novelty and often perform poorly in partially observable settings. Moreover, they typically ignore the role of actions in driving observation changes, as not all actions lead to meaningful state transitions. To overcome these limitations, we propose a trajectory-level novelty measure that estimates the novelty of a state by comparing current observations with past ones along the trajectory. To focus on meaningful exploration, we incorporate the mutual information between actions and trajectory novelty to filter out random fluctuations and retain only novelty caused by the agent's actions. Additionally, we introduce a first-visit constraint on observation–action pairs, rewarding only interactions that result in state transitions to enhance exploration efficiency. We conducted experiments in the MiniGrid-ObstructedMaze environment characterised by complex object interactions and sparse rewards. Results demonstrate that our method achieves state-of-the-art performance in convergence speed and average returns. Furthermore, it shows strong generalisation on high-dimensional Atari benchmarks and demonstrates robust performance in more challenging MiniGrid variants. Implementation code is available at: https://github.com/MurrayMa0816/TNCOA.

With the development of unmanned aerial vehicle and satellite technology, the application of tiny object detection in remote sensing images is becoming increasingly widespread. Although significant progress has been made in the accuracy and speed of object detection in recent years, performance declines sharply when general object detectors are applied to tiny objects; one of the main reasons is unsuitable label assignment strategy. Traditional label assignment strategies often rely on fixed thresholds, leading to mismatches between the number of positive samples and object areas. Additionally, most improved methods require setting one or more hyperparameters. In this paper, we propose a dynamic adaptive label assignment strategy (DALA) comprising three modules. First, we calculate the similarity distance to comprehensively evaluate the matching degree between anchors and each ground truth. Then, we use the ratio-based label assignment strategy to select an appropriate number of positive samples for each object. Finally, we introduce dynamic weighting loss during training to ensure the model pays more attention to tiny objects. Our three modules automatically adapt to different datasets and detectors without any manual hyperparameter settings. Extensive experiments on four widely used datasets demonstrate the excellent performance of our proposed method. Our code will be released soon.

Intelligent and adaptive defence systems that can quickly thwart changing cyberthreats are becoming more and more necessary in the dynamic and data-intensive Internet of things (IoT) environment. Using the NSL-KDD benchmark dataset, this paper presents an improved anomaly detection system that combines an optimised sequential neural network (OSNN) with an XGBoost model optimised using Bayesian approaches. Key drawbacks of conventional intrusion detection systems (IDSs), including manual hyperparameter tuning, extensive preprocessing and ineffective optimisation techniques, are successfully addressed by the suggested solution. By automating the adjustment of key parameters, such as learning rate, tree depth and regularisation strength, Bayesian optimisation enhances model convergence, prediction accuracy and generalisation. Superior binary classification performance was attained by the optimised XGBoost, which had 99.98% accuracy, 99.94% F1 score and 99.95% MCC. In a similar vein, the Bayesian-optimised OSNN obtained flawless precision and recall for the normal and U2R classes, achieving 99.92% overall accuracy in multiclass identification. Across all attack types, its CNN-based architecture showed excellent sensitivity and specificity, attaining almost flawless AUC values. The suggested Bayesian-optimised framework offers a highly accurate, scalable and self-adaptive intrusion detection solution for actual IoT network environments, outperforming traditional machine and deep learning models by a large margin overall.

Transformers have been widely applied to hyperspectral image classification, leveraging their self-attention mechanism for powerful global modelling. However, two key challenges remain as follows: excessive memory and computational costs from calculating correlations between all tokens (especially as image size or spectral bands increase) and limited ability to model local boundary information due to lacking explicit enhancement mechanisms. This paper proposes a novel method, bridge transformer network fused with deep graph convolution (BTDGC), to address these issues. The framework includes three components as follows: a double random masking mechanism (DRMM) that forces the model to infer masked features from context during training, a bridge transformer (BT) module with bridge tokens for cross-region feature interaction and a Deep Graph Convolutional Pooling (DGCP) module that preserves spatial topology while aggregating hierarchical information. Experiments on standard hyperspectral datasets show BTDGC outperforms mainstream methods in classification accuracy and robustness, effectively balancing global modelling and local boundary representation. The code is available at https://github.com/jenny3489/BTDGC.

As a government-regulated public service, traffic signal control (TSC) requires reliable and transparent decision-making. However, existing deep reinforcement learning (DRL) methods, despite improvements in control accuracy, still lack explainability and generalisation, severely limiting their applicability in real-world environments. To address the challenges above, this paper proposes GenEx-TSC, a generalisable and explainable TSC method that integrates deep reinforcement learning with large language models (LLMs). First, starting from vehicle-level states, we train a DRL agent incorporating intersection physical heterogeneity and neighbourhood information, which lays the evaluation foundation for constructing a high-quality LLM dataset. Subsequently, the LLM agent is optimised through a two-stage training mechanism. In the distillation stage, a lightweight LLM agent is trained using the reasoning trajectories of a larger-scale LLM agent, inheriting its semantic understanding and decision-generation capabilities and in the alignment stage, the DRL evaluation network is employed to calibrate the outputs of the distilled LLM agent, ensuring that the generated cycle-level signal timing strategies are both efficient and interpretable. We synthesise 10 intersection networks with different physical attributes in SUMO and set traffic flows of varying scales. Experimental results across diverse traffic environments demonstrate that the proposed GenEx-TSC exhibits clear advantages over traditional methods, mainstream DRL methods and LLM baselines in terms of control accuracy, generalisation and explainability.

A formation inversion algorithm with real-time performance and accuracy is crucial for natural gamma logging while drilling (LWD). However, traditional inversion algorithms are often limited by high computational resource consumption and insufficient accuracy. To address these issues, an improved forward method for natural gamma LWD is proposed. The inverse problem is subsequently modelled using the proposed forward method through which the search methodology and region of formation information are determined. On this basis, a collaborative fuzzy gradient neural dynamics (CFGND) algorithm is proposed, which combines the advantages of the collaborative mechanism in swarm intelligence algorithms and fuzzy gradient neural dynamics (FGND) to improve its accuracy and real-time performance. Specifically, the collaborative mechanism is applied to conduct a global search using all possible formation information. Concurrently, the FGND algorithm initiates a local search from each particle and dynamically and intelligently adjusts the learning rate of the neural dynamics through a fuzzy logic system during the process to achieve rapid and stable local convergence. The CFGND algorithm subsequently updates its globally optimal solution using the optimal solution obtained from the FGND algorithm. This iterative process continues until the termination condition is met. Theoretical analysis proves the existence of an optimal solution for the inverse problem and the convergence of the CFGND algorithm. The results of simulations and experiments demonstrate that the proposed formation inversion algorithm features high accuracy and sufficient real-time performance.

Analysing learners' facial expressions during learning and exploring their learning processes and emotional changes are of great significance for assisting teachers' teaching and promoting smart education. In complex learning environments, static facial expression recognition fails to capture the dynamic changes of learners' expressions losing the continuous features in the learning process, and its recognition effect is easily interfered with by factors such as occlusion and lighting variations during learning. To address the above issues, a network model based on adaptive global attention and temporal difference is proposed to recognise learners' dynamic expression sequences. Firstly, we have designed an Adaptive Global Attention (AGA) block, which adaptively models inter-channel relationships to dynamically enhance key channels that are highly correlated with learners' states while suppressing redundant information, thereby improving the model's feature representation capability under noisy environments. Secondly, we have designed a Differential Temporal Transformer (DTFormer) to extract differential information between consecutive frames, increasing the model's sensitivity to learners' facial expression dynamics and improving recognition performance. The two components complement each other in terms of spatial feature enhancement and temporal dynamic modelling effectively improving the model's overall capability for representing learners' dynamic facial expressions. Experiments were conducted on public datasets DFEW, FERV39k and the learner E-learning emotional state data set DAiSEE, and comparisons were made with classical methods using objective indicators. The results demonstrate that the proposed method outperforms the comparison methods in multiple performance indicators, thereby verifying its effectiveness.

This paper presents an event-triggered pinning observer-based control of unknown large-scale interconnected systems under false data injection attacks using distributed noisy data without any system identification step. An exogenous system is used to model the false data injection attack. Two event-triggered mechanisms are introduced at the sensor and observer sides to reduce unnecessary computational resource consumption. The corresponding data representations are constructed for unknown systems subject to attacks. Under the framework of the data representation of unknown systems, an event-triggered pinning extended observer-based controller is designed. The global exponential stability of unknown systems is reached under the designed controller. Finally, an example of non-isothermal continuous stirred tank reactors is provided to verify the effectiveness of the obtained results.

The analysis of spatiotemporal data is essential across many fields, such as transportation, meteorology and healthcare. Data gathered in practical applications often suffer from incompleteness due to device failures and network disruptions. Spatiotemporal imputation targets the estimation of missing observations by exploiting intrinsic spatial–temporal dependencies. Although traditional statistical and machine-learning methods depend on restrictive distributional assumptions, graph- or recurrent-based models accumulate errors through iterative propagation. Diffusion probabilistic models mitigate these issues by sampling directly from a learnt data prior instead of recycling past imputations. However, existing conditional diffusion variants still converge towards overly similar reconstructions, obscuring the genuine uncertainty and heterogeneity of real-world traffic, environmental or clinical streams. Preserving-and faithfully quantifying-this intrinsic diversity is crucial for reliable forecasting and downstream decision-making. We propose C²TSD, a conditional diffusion framework that integrates disentangled temporal representations and contrastive learning to improve generalisability in spatiotemporal imputation. Specifically, the approach uses disentangled temporal representations as conditional information to guide the reverse process. We also enhance the final loss using a contrastive learning strategy to improve representation quality, mitigating the impact of data missing completely at random (MCAR) and noise on learnt features. Through comprehensive experiments using three distinct real-world datasets, C²TSD has competitive results compared to leading-edge baselines.

Object detection of unmanned firefighting vehicles faces challenges such as strong electromagnetic interference, drastic lighting changes and dynamic object variations. To address these issues, we propose a two-stage 3D point cloud object detection algorithm called TED-CasA-Fusion. The first stage uses the transformation-equivariant detector backbone that explicitly models rotation/reflection equivariance via weight-sharing sparse convolutions, which improves detection robustness to dynamically transformed objects. The second stage introduces a cascade attention-based multistage refinement network that aggregates cross-stage object features through cascade attention modules, which effectively enhances feature representation for multiscale objects. Furthermore, the second stage integrates weighted bounding box voting to address training imbalance due to dense nearby and sparse distant point distributions, thereby improving detection accuracy for distant and sparse targets. Comparative experiments were conducted on the KITTI dataset and a self-collected firefighting dataset between the proposed algorithm and some state-of-the-art algorithms. Results show that the proposed algorithm achieves the best 3D detection accuracy for hard-category objects on the KITTI dataset and also outperforms other detection approaches on the firefighting dataset. This work offers an efficient and reliable solution to environmental perception of unmanned firefighting vehicles.

Integrating healthcare systems with intelligent transportation networks represents a critical frontier in modern urban infrastructure, where efficient resource allocation and timely service delivery can significantly impact patient outcomes. However, current approaches often fail to capture the complex interplay between healthcare facility accessibility and transportation dynamics, particularly during emergencies. Additionally, the temporal dependencies in healthcare service delivery follow strict sequential patterns that significantly influence both routine operations and emergency response effectiveness. To address these challenges, we propose a multi-scale spatio-temporal transformer network for healthcare and transportation (MST-HT) that leverages generative AI capabilities. Our model employs multiple specialised transformer networks to model different spatial scales, capturing hidden dependencies while using graph convolutional networks to learn static infrastructure features. The architecture incorporates healthcare district patterns, emergency response corridors and facility distributions through a novel gating mechanism that adaptively combines features based on their predictive importance. The model maintains awareness of critical service delivery patterns by embedding healthcare-specific temporal position information while optimising resource allocation. Experiments on real-world datasets demonstrate MST-HT's superior performance, achieving a 15.7% reduction in emergency response times and a 23.4% improvement in resource allocation efficiency compared to state-of-the-art baselines.

With the advancement of satellite remote sensing technology, object detection based on high-resolution remote sensing imagery has emerged as a prominent research focus in the field of computer vision. Although numerous algorithms have been developed for remote sensing image object detection, they still suffer from challenges such as low detection accuracy and high false positive rates. To address these issues, we propose a novel architecture, the multiscale feature fusion network (MSFFNet). MSFFNet is composed of three key components: the Large Selective Kernel Block (LSKBlock), the Space-to-Depth ADown (SPDA) module and the Double Feature Aggregation Neck (DFAN). Specifically, the LSKBlock adaptively captures salient target features by dynamically adjusting the receptive field size, thereby enhancing detection precision. The SPDA module converts spatial correlations into channel-wise dependencies by segmenting and reordering the feature maps, which helps preserve fine-grained information, suppress background interference and reduce false detections. Furthermore, the DFAN integrates shallow and deep features through a multiscale feature fusion module (MSFFM), enabling the extraction of multiscale target representations and improving overall detection performance. Extensive experiments on public datasets, SIMD, VisDrone2019 and DIOR, demonstrate the effectiveness of our approach. Compared with the YOLOv9s baseline model, MSFFNet achieves improvements in mAP50% of 0.6%, 1.9% and 3.5%, respectively.

Graph contrastive learning (GCL) has emerged as a dominant paradigm for self-supervised representation learning for attributed graph data. However, existing GCL methods heavily rely on empirical graph data augmentation, which may distort intrinsic graph semantics and produce poor generalisation without carefully chosen or designed augmentation techniques. Furthermore, most GCL approaches focus on same-granularity contrastive learning (e.g., node vs. node), neglecting the hierarchical and multigranular properties inherent in real-world networks, leading to suboptimal performance. To address these limitations, we propose HPoolGCL, a cross-granularity GCL framework compatible with various hierarchical graph pooling methods to capture multigranularity information. Our framework eliminates the need for handcrafted augmentations, explicit negative sampling and complex multiencoder architectures by applying two novel loss functions in hierarchical graph pooling. The theoretical analysis is provided to explain the effectiveness of unified MGC and HiCR losses from three perspectives, namely, the information maximisation principle, the redundancy reduction principle and the information bottleneck principle. The experimental results demonstrate that HPoolGCL achieves state-of-the-art performance across multiple downstream tasks on five benchmarks. Our codes are available at https://github.com/Heycen/HPoolGCL.