Training Convergence Research Articles

Dynamical transition in controllable quantum neural networks with large depth

Understanding the training dynamics of quantum neural networks is a fundamental task in quantum information science with wide impact in physics, chemistry and machine learning. In this work, we show that the late-time training dynamics of quantum neural networks with a quadratic loss function can be described by the generalized Lotka-Volterra equations, leading to a transcritical bifurcation transition in the dynamics. When the targeted value of loss function crosses the minimum achievable value from above to below, the dynamics evolve from a frozen-kernel dynamics to a frozen-error dynamics, showing a duality between the quantum neural tangent kernel and the total error. In both regions, the convergence towards the fixed point is exponential, while at the critical point becomes polynomial. We provide a non-perturbative analytical theory to explain the transition via a restricted Haar ensemble at late time, when the output state approaches the steady state. Via mapping the Hessian to an effective Hamiltonian, we also identify a linearly vanishing gap at the transition point. Compared with the linear loss function, we show that a quadratic loss function within the frozen-error dynamics enables a speedup in the training convergence. The theory findings are verified experimentally on IBM quantum devices.

A Joint Estimation Method of Distribution Network Topology and Line Parameters Based on Power Flow Graph Convolutional Networks

Accurate identification of network topology and line parameters is essential for effective management of distribution systems. An innovative joint estimation method for distribution network topology and line parameters is presented, utilizing a power flow graph convolutional network (PFGCN). This approach addresses the limitations of traditional methods that rely on costly voltage phase angle measurements. The node correlation principle is applied to construct a node correlation matrix, and a minimum distance iteration algorithm is proposed to generate candidate topologies, which serve as graph inputs for the parameter estimation model. Based on the topological dependencies and convolutional properties of AC power flow equations, a PFGCN model is designed for line parameter estimation. Parameter refinement is achieved through an alternating iterative process of pseudo-trend calculation and neural network training. Training convergence and loss function values are used as feedback to filter and validate candidate topologies, enabling precise joint estimation of both topologies and parameters. The proposed method’s accuracy, transferability, and robustness are demonstrated through experiments on the IEEE-33 and modified IEEE-69 distribution systems. Multiple metrics, including MAPE, IAE, MAE, and R2, highlight the proposed method’s advantages over Adaptive Ridge Regression (ARR). In the C33 scenario, the proposed method achieves MAPEs of 4.6% for g and 5.7% for b, outperforming the ARR method with MAPEs of 7.1% and 7.9%, respectively. Similarly, in the IC69 scenario, the proposed method records MAPEs of 3.0% for g and 5.9% for b, surpassing the ARR method’s 5.1% and 8.3%.

Audio feature enhancement based on quaternion filtering and deep hashing

This paper aims to solve the problems of difficult convergence of audio model training, large data demand, and large dimensionality of storage space for audio-generated feature vectors. To this end, this paper proposes the use of quaternion Gabor filtering to suppress the background information of the spectrogram and reduce the interference of the data for the case of insufficient data alignment between audio data and image data after shifting the domain. In addition, different scales of window lengths and frame shifts are used to capture the connections between different vocal objects. To address the problem that the generated feature vectors are large dimensional, we use a deep hash module to map high-dimensional features to low-dimensional features and use a probability function to make the learned samples more consistent with the overall distribution. In the experimental evaluation, the proposed method was evaluated on the environmental sound classification dataset and the music genre classification dataset. The proposed method uses only a common backbone network and improves the accuracy of audio recognition.

ETLSH-YOLO: An Edge–Real-Time Transmission Line Safety Hazard Detection Method

Using deep learning methods to detect potential safety hazards in transmission lines is the mainstream method for power grid security monitoring. However, the existing model is too complex to adapt to edge device deployment and real-time detection. Therefore, an edge–real-time transmission line safety hazard detection method (ETLSH-YOLO) was proposed to reduce the model’s complexity and improve the model’s robustness. Firstly, a re-parameterized Ghost efficient layer aggregation network (RepGhostCSPELAN) was designed to effectively fuse the feature information of different layers while enhancing the model’s expression ability and reducing the number of model parameters and floating-point operations. Then, a spatial channel decoupled downsampling block (CSDovn) was designed to reduce computational redundancy and improve the computational efficiency of the model. Then, coordinate attention (CA) was added in the process of multi-scale feature fusion to suppress the interference of complex background and improve the global perception ability of the model object. Finally, the Mish activation function was used to improve the network’s training speed, convergence, and generalization ability. The experimental results show that the mAP50 of this model improved by 1.73% compared with the baseline model, and the number of parameters and floating-point operations were reduced by 33.96% and 22.22%, respectively. This model lays the foundation for solving the dilemma of edge device deployment.

Batch channel normalized-CWGAN with Swin Transformer for imbalanced data fault diagnosis of rotating machinery

Abstract In real scenarios, rotating machinery is mainly operated in optimal condition, leading to fault data scarce and difficult to collect. This issue results in imbalanced data, significantly limiting the effectiveness of intelligent fault diagnosis methods. To address this issue, a novel fault diagnosis method for rotating machinery is proposed in this paper, which combines the batch channel normalized conditional wasserstein generative adversarial network (BCN-CWGAN) with Swin Transformer. Firstly, the one-dimensional vibration signal is preprocessed into two-dimensional feature images using a symmetrized dot pattern (SDP). Subsequently, self-attention mechanism and deep feature learning module constructed by DenseNet are integrated into the generator of GAN to acquire more discriminative feature information. Meanwhile, the discriminator of GAN is combined with batch channel normalization strategy, which further enhances the generalization ability. Besides, a two time-scale update rule strategy enhances training stability and convergence speed by updating model parameters at different time scales. Then, the data augmentation capability of BCN-CWGAN is used to generate high-quality fault samples to augment the imbalanced dataset. Finally, Swin Transformer is combined to achieve accurate fault diagnosis. The performance enhancement of the proposed method is verified through comparison and diagnosis results of two engineering experiments, demonstrating its substantial value for research in engineering practice. With the proposed data augmentation method, the average accuracy of A1, B1, C1, and D1 datasets in experiment 1 reached 99.24%, 98.85%, 96.78%, and 96.04%, respectively. Meanwhile, the proposed method achieved the best accuracy in experiment 2.

Quantum long short-term memory (QLSTM) vs. classical LSTM in time series forecasting: a comparative study in solar power forecasting

Accurate solar power forecasting is pivotal for the global transition towards sustainable energy systems. This study conducts a meticulous comparison between Quantum Long Short-Term Memory (QLSTM) and classical Long Short-Term Memory (LSTM) models for solar power production forecasting. The primary objective is to evaluate the potential advantages of QLSTMs, leveraging their exponential representational capabilities, in capturing the intricate spatiotemporal patterns inherent in renewable energy data. Through controlled experiments on real-world photovoltaic datasets, our findings reveal promising improvements offered by QLSTMs, including accelerated training convergence and substantially reduced test loss within the initial epoch compared to classical LSTMs. These empirical results demonstrate QLSTM’s potential to swiftly assimilate complex time series relationships, enabled by quantum phenomena like superposition. However, realizing QLSTM’s full capabilities necessitates further research into model validation across diverse conditions, systematic hyperparameter optimization, hardware noise resilience, and applications to correlated renewable forecasting problems. With continued progress, quantum machine learning can offer a paradigm shift in renewable energy time series prediction, potentially ushering in an era of unprecedented accuracy and reliability in solar power forecasting worldwide. This pioneering work provides initial evidence substantiating quantum advantages over classical LSTM models while acknowledging present limitations. Through rigorous benchmarking grounded in real-world data, our study illustrates a promising trajectory for quantum learning in renewable forecasting.

Mesh-Free Solution of 2D Poisson Equation with High Frequency Charge Patterns Using Data-Free Physics Informed Neural Network

Abstract In this paper, we present a data-free physics-informed neural networks (PINNs) approach for solving two-dimensional (2D) Poisson equation, which is pivotal in fields such as electromagnetics, mechanical engginering, and thermodynamics. Traditional numerical method for solving this equation often require structured mesh generation such as Finite Element Method (FEM), which can be computationally expensive when dealing with high resolution Poisson Equation Solution. To address this challenge, we leverage the capabilities of PINNs capturing pattern of complex system by incorporating physical law and boundary condition as part of loss function on training model. While PINNs provide a robust framework for solving differential equations within boundary condition, they have struggle with capturing high-frequency pattern due to smooth nature of typical activation function used in neural networks. To evercome this issue, we enhance our model by incorporating Fourier Features Networks, which map inputs through a series of sinusoidal functions before feeding the input into the neural network. The result show that Fourier feature network can enhance convergence of training of PINNs model faster and obtained better result than PINNs without Fourier feature networks.

Prediction of Crohn's disease based on deep feature recognition

BackgroundCrohn's disease is a complex genetic disease that involves chronic gastrointestinal inflammation and results from a complex set of genetic, environmental, and immunological factors. By analyzing data from the human microbiome, genetic information can be used to predict Crohn's disease. Recent advances in deep learning have demonstrated its effectiveness in feature extraction and the use of deep learning to decode genetic information for disease prediction. MethodsIn this paper, we present a deep learning-based model that utilizes a sequential convolutional attention network (SCAN) for feature extraction, incorporates adaptive additive interval losses to enhance these features, and employs support vector machines (SVM) for classification. To address the challenge of unbalanced Crohn's disease samples, we propose a random noise one-hot encoding data augmentation method. ResultsData augmentation with random noise accelerates training convergence, while SCAN-SVM effectively extracts features with adaptive additive interval loss enhancing differentiation. Our approach outperforms benchmark methods, achieving an average accuracy of 0.80 and a kappa value of 0.76, and we validate the effectiveness of feature enhancement. ConclusionsIn summary, we use deep feature recognition to effectively analyze the potential information in genes, which has a good application potential for gene analysis and prediction of Crohn's disease.

Real-Time Identification of Strawberry Pests and Diseases Using an Improved YOLOv8 Algorithm

Strawberry crops are susceptible to a wide range of pests and diseases, some of which are insidious and diverse due to the shortness of strawberry plants, and they pose significant challenges to accurate detection. Although deep learning-based techniques to detect crop pests and diseases are effective in addressing these challenges, determining how to find the optimal balance between accuracy, speed, and computation remains a key issue for real-time detection. In this paper, we propose a series of improved algorithms based on the YOLOv8 model for strawberry disease detection. These include improvements to the Convolutional Block Attention Module (CBAM), Super-Lightweight Dynamic Upsampling Operator (DySample), and Omni-Dimensional Dynamic Convolution (ODConv). In experiments, the accuracy of these methods reached 97.519%, 98.028%, and 95.363%, respectively, and the F1 evaluation values reached 96.852%, 97.086%, and 95.181%, demonstrating significant improvement compared to the original YOLOv8 model. Among the three improvements, the improved model based on CBAM has the best performance in training stability and convergence, and the change in each index is relatively smooth. The model is accelerated by TensorRT, which achieves fast inference through highly optimized GPU computation, improving the real-time identification of strawberry diseases. The model has been deployed in the cloud, and the developed client can be accessed by calling the API. The feasibility and effectiveness of the system have been verified, providing an important reference for the intelligent research and application of strawberry disease identification.

Curriculum Design and Sim2Real Transfer for Reinforcement Learning in Robotic Dual-Arm Assembly

Robotic systems are crucial in modern manufacturing. Complex assembly tasks require the collaboration of multiple robots. Their orchestration is challenging due to tight tolerances and precision requirements. In this work, we set up two Franka Panda robots performing a peg-in-hole insertion task of 1 mm clearance. We structure the control system hierarchically, planning the robots’ feedback-based trajectories with a central policy trained with reinforcement learning. These trajectories are executed by a low-level impedance controller on each robot. To enhance training convergence, we use reverse curriculum learning, novel for such a two-armed control task, iteratively structured with a minimum requirements and fine-tuning phase. We incorporate domain randomization, varying initial joint configurations of the task for generalization of the applicability. After training, we test the system in a simulation, discovering the impact of curriculum parameters on the emerging process time and its variance. Finally, we transfer the trained model to the real-world, resulting in a small decrease in task duration. Comparing our approach to classical path planning and control shows a decrease in process time, but higher robustness towards calibration errors.

An automatic method for scoring poultry footpad dermatitis with deep learning and thermal imaging

Footpad dermatitis (FPD) is a common poultry condition that can negatively influence chickens’ production, welfare, and health. However, no automated tool for monitoring FPD in live chickens is currently available. The objective of this study was to develop and optimize deep learning models to monitor hens’ FPD scores (i.e., 0–2 scale with higher scores indicating poorer footpad conditions). A total of 700 Hy-Line W-36 hens were raised in four cage-free housing systems integrated with Electrostatic Particle Ionization and various bedding materials. A GoPro camera with an upward lens was placed inside a transparent box. Individual laying hens were placed on the top surface of the box to acquire RGB images. In addition, a thermal camera was used to record RGB and thermal images of footpads, and the images were manually scored to assess their footpad conditions. Preprocessing techniques (e.g., filtration, separation, and augmentation) were deployed to enhance dataset quality and size. Moreover, YOLOv8 models (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x) and YOLOv7 models (YOLOv7 and YOLOv7x) were comparatively evaluated for predicting FPD scores. The results show that the YOLOv8l outperformed other models, with higher recall (96.6 %), mAP@0.50 (97.0 %), and F1-score (95.0 %). Additionally, the YOLOv8l-FPD model exhibited a high mAP@0.50 for score 0 (98.0 %), score 1 (95.0 %), and score 2 (97.9 %) and F1-score (95.0 %) for all FPD scores. Notably, using thermal images could result in faster convergence of model training and slightly better FPD score prediction performance than RGB images. The proposed technique can be useful for non-invasive automatic FPD scoring and further improve automation levels and animal welfare in the egg industry.

Enhancing Clinical Accuracy of Medical Chatbots with Large Language Models.

The rapid advancement of large language models (LLMs) has opened up new possibilities for transforming healthcare practices, patient interactions, and medical report generation. This paper explores the application of LLMs in developing medical chatbots and virtual assistants that prioritize clinical accuracy. We propose a novel multi-turn dialogue model, including adjusting the position of layer normalization to improve training stability and convergence, employing a contextual sliding window reply prediction task to capture fine-grained local context, and developing a local critical information distillation mechanism to extract and emphasize the most relevant information. These components are integrated into a multi-turn dialogue model that generates coherent and clinically accurate responses. Experiments on the MIMIC-III and n2c2 datasets demonstrate the superiority of the proposed model over state-of-the-art baselines, achieving significant improvements in perplexity, BLEU-2, recall at K scores, medical entity recognition, and response coherence. The proposed model represents a significant step in developing reliable and contextually relevant multi-turn medical dialogue systems that can assist patients and healthcare professionals.

Three-dimensional cooperative guidance with impact angle constraints via value-policy decomposed multi-agent reinforcement learning

For a three-dimensional impact angle-constrained time-coordination guidance problem, a value-policy decomposed multi-agent twin delayed deep deterministic (VPD-MATD3) policy gradient algorithm is proposed in this paper, and a temporal-spatial cooperative guidance law is designed consequently. The derived cooperative engagement kinematics is formulated as a value-policy decomposed partially observable Markov decision process. Value decomposition and policy decomposition are correspondingly proposed to address the two bottlenecks of reward coupling and policy coupling in 3D guidance, thus improving the training process. The training environment with multiple uncertainties, such as observation noise, maneuverability perturbation, autopilot time lag, switching communication topology, and random initialization, is subsequently constructed. Time-energy type reward functions are designed, incorporating energy consumption as a direct optimization indicator via a penalty term. A long short-term memory network layer is inserted inside the policy networks to suppress the negative effect of observation noise. The policy training curves verify the significant effect of the proposed algorithm in improving training convergence. The policy testing results illustrate the robustness and generalization of the VPD-MATD3 cooperative guidance law and its ability to cope with randomly switched communication topologies. Moreover, the comparative testing further reveals its time-energy suboptimal property.

Translation Consistent Semi-supervised Segmentation for 3D Medical Images.

3D medical image segmentation methods have been successful, but their dependence on large amounts of voxel-level annotated data is a disadvantage that needs to be addressed given the high cost to obtain such annotation. Semi-supervised learning (SSL) solves this issue by training models with a large unlabelled and a small labelled dataset. The most successful SSL approaches are based on consistency learning that minimises the distance between model responses obtained from perturbed views of the unlabelled data. These perturbations usually keep the spatial input context between views fairly consistent, which may cause the model to learn segmentation patterns from the spatial input contexts instead of the foreground objects. In this paper, we introduce the Translation Consistent Co-training (TraCoCo) which is a consistency learning SSL method that perturbs the input data views by varying their spatial input context, allowing the model to learn segmentation patterns from foreground objects. Furthermore, we propose a new Confident Regional Cross entropy (CRC) loss, which improves training convergence and keeps the robustness to co-training pseudo-labelling mistakes. Our method yields state-of-the-art (SOTA) results for several 3D data benchmarks, such as the Left Atrium (LA), Pancreas-CT (Pancreas), and Brain Tumor Segmentation (BraTS19). Our method also attains best results on a 2D-slice benchmark, namely the Automated Cardiac Diagnosis Challenge (ACDC), further demonstrating its effectiveness. Our code, training logs and checkpoints are available at https://github.com/yyliu01/ TraCoCo.

Integrated Intelligent Control of Redundant Degrees-of-Freedom Manipulators via the Fusion of Deep Reinforcement Learning and Forward Kinematics Models

Redundant degree-of-freedom (DOF) manipulators offer increased flexibility and are better suited for obstacle avoidance, yet precise control of these systems remains a significant challenge. This paper addresses the issues of slow training convergence and suboptimal stability that plague current deep reinforcement learning (DRL)-based control strategies for redundant DOF manipulators. We propose a novel DRL-based intelligent control strategy, FK-DRL, which integrates the manipulator’s forward kinematics (FK) model into the control framework. Initially, we conceptualize the control task as a Markov decision process (MDP) and construct the FK model for the manipulator. Subsequently, we expound on the integration principles and training procedures for amalgamating the FK model with existing DRL algorithms. Our experimental analysis, applied to 7-DOF and 4-DOF manipulators in simulated and real-world environments, evaluates the FK-DRL strategy’s performance. The results indicate that compared to classical DRL algorithms, the FK-DDPG, FK-TD3, and FK-SAC algorithms improved the success rates of intelligent control tasks for the 7-DOF manipulator by 21%, 87%, and 64%, respectively, and the training convergence speeds increased by 21%, 18%, and 68%, respectively. These outcomes validate the proposed algorithm’s effectiveness and advantages in redundant manipulator control using DRL and FK models.

Method for Remaining Useful Life Prediction of Turbofan Engines Combining Adam Optimization-Based Self-Attention Mechanism with Temporal Convolutional Networks

Conducting the remaining useful life (RUL) prediction for an aircraft engines is of significant importance in enhancing aircraft operation safety and formulating reasonable maintenance plans. Addressing the issue of low prediction model accuracy due to traditional neural networks’ inability to fully extract key features, this paper proposes an engine RUL prediction model based on the adaptive moment estimation (Adam) optimized self-attention mechanism–temporal convolutional network (SAM-TCN) neural network. Firstly, the raw data monitored by sensors are normalized, and RUL labels are set. A sliding window is utilized for overlapping sampling of the data, capturing more temporal features while eliminating data dimensionality. Secondly, the SAM-TCN neural network prediction model is constructed. The temporal convolutional network (TCN) neural network is used to capture the temporal dependency between data, solving the mapping relationship of engine degradation characteristics. A self-attention mechanism (SAM) is employed to adaptively assign different weight contributions to different input features. In the experiments, the root mean square error (RMSE) values on four datasets are 11.50, 16.45, 11.62, and 15.47 respectively. These values indicate further reduction in errors compared to methods reported in other literature. Finally, the SAM-TCN prediction model is optimized using the Adam optimizer to improve the training effectiveness and convergence speed of the model. Experimental results demonstrate that the proposed method can effectively learn feature data, with prediction accuracy superior to other models.

Power quality optimization technology for photovoltaic nanoscale electronic grid-connected systems based on FNN algorithm

With the continuous maturity and widespread application of photovoltaic power generation technology, the power quality of photovoltaic grid-connected systems has become a significant factor affecting the power grid’s stability and the lifespan of equipment. This research aims to improve the power quality in photovoltaic nanoscale electronic grid-connected systems. Through in-depth analysis of harmonic problems in grid-connected systems, a harmonic prediction and governance means using a fuzzy neural network algorithm was put forward. To improve the accuracy and adaptability of predictions, the long short-term memory network was combined with a fuzzy neural network to form an improved deep learning model. Comparative experiments confirmed that the research prediction model exhibited advantages in training convergence, prediction accuracy, and various error evaluation indicators, with an accuracy of up to 96.24% and an R2 value as high as 0.9998. It effectively reduced harmonic content by 2.4% and significantly improved the harmonic compensation effect. Based on the above-mentioned model, dynamic compensation governance for harmonics was implemented. By predicting the harmonic content in grid-connected electricity in real-time and taking corresponding measures for compensation, the adverse effects of harmonics on the entire grid-connected system were effectively reduced. The deep learning-based power quality optimization technology developed in this research can accurately predict harmonic problems in photovoltaic nanoscale electronic grid-connected systems. Moreover, it can provide efficient compensation governance solutions and help improve the entire power grid’s energy efficiency and operational stability.

Enhancing Temporal Knowledge Graph Representation with Curriculum Learning

Temporal knowledge graph representation approaches encounter significant challenges in handling the complex dynamic relations among entities, relations, and time. These challenges include the high difficulty of training and poor generalization performance, particularly with large datasets. To address these issues, this paper introduces curriculum learning strategies from machine learning, aiming to improve learning efficiency through effective curriculum planning. The proposed framework constructs a high-dimensional filtering model based on graph-based high-order receptive fields and employs a scoring model that uses a curriculum temperature strategy to evaluate the difficulty of temporal knowledge graph data quadruples at each stage. By progressively expanding the receptive field and dynamically adjusting the difficulty of learning samples, the model can better understand and capture multi-level information within the graph structure, thereby improving its generalization capabilities. Additionally, a temperature factor is introduced during model training to optimize parameter gradients, alongside a gradually increasing training strategy to reduce training difficulty. Experiments on the benchmark datasets ICEWS14 and ICEWS05-15 demonstrate that this framework not only significantly enhances model performance on these datasets but also substantially reduces training convergence time.

Enhancing YOLO for occluded vehicle detection with grouped orthogonal attention and dense object repulsion

In real-life complex traffic environments, vehicles are often occluded by extraneous background objects and other vehicles, leading to severe degradation of object detector performance. To address this issue, we propose a method named YOLO-OVD (YOLO for occluded vehicle detection) and a dataset for effectively handling vehicle occlusion in various scenarios. To highlight the model attention in unobstructed region of vehicles, we design a novel grouped orthogonal attention (GOA) module to achieve maximum information extraction between channels. We utilize grouping and channel shuffling to address the initialization and computational issues of original orthogonal filters, followed by spatial attention for enhancing spatial features in vehicle-visible regions. We introduce a CIoU-based repulsion term into the loss function to augment the network’s localization accuracy in scenarios involving densely packed vehicles. Moreover, we explore the effect of the knowledge-based Laplacian Pyramid on the OVD performance, which contributes to fast convergence in training and ensures more detailed and comprehensive feature retention. We conduct extensive experiments on the established occluded vehicle detection dataset, which demonstrates that the proposed YOLO-OVD model significantly outperforms 14 representative object detectors. Notably, it achieves improvements of 4.7% in Precision, 3.6% in AP@0.5, and 1.9% in AP@0.5:0.95 compared to the YOLOv5 baseline.

Review on Temporal Convolutional Networks for Electricity Theft Detection with Limited Data

Electricity theft detection using artificial intelligence (AI) and machine learning techniques have shown significant promise in recent research. However, practical implementation and widespread adoption of these advanced methods face several persistent challenges, particularly when dealing with limited data. This review delves into the computational complexity, data requirements, overfitting issues, and scalability and generalizability concerns associated with popular techniques such as Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM), Deep Convolutional Neural Networks (DCNN), Multi-Layer Perceptron (MLP), Gated Recurrent Unit (GRU), and Artificial Neural Networks (ANN). Computational complexity and resource constraints affect the training times and convergence of TCN, LSTM, and DCNN, while high data needs and parameter tuning hinder MLP and GRU. The ANN-based method utilized by the Electricity Company of Ghana underscores overfitting and data duplication, further exacerbated by limited data availability. Moreover, the scalability and generalizability of TCN, LSTM, and DCNN across different regions and larger datasets are limited, with effectiveness varying based on electricity consumption patterns and theft tactics. Addressing these challenges through optimizing computational efficiency, improving data quality and utilization, and enhancing scalability and generalizability is crucial, especially in data-constrained environments. Continued research and development in these areas will be essential for realizing the full potential of AI-based electricity theft detection systems with limited data.