Applied Intelligence最新文献

A novel Reinforcement Learning (RL) approach Multi-agent Joint Control with State-Action Embedding (MASAE) is proposed in this paper to address the sample-efficiency issue of RL in robot control. It combines the relative entropy regularization and high update-to-data (UTD) ratios in one multi-agent framework to accelerate the learning process while naturally mitigating the overestimation of value functions caused by high UTD ratios by multiple agents. The state-action embeddings are employed to adaptively abstract the hidden features behind the state-action space for enhanced learning efficiency. Evaluated by several simulated benchmark control tasks and a real-world Unitree Go1 quadruped robot system, MASAE demonstrates significant advantages in learning capability and sampling efficiency compared to various related RL baselines, indicating its potential in learning challenging real-world robot systems with a limited number of samples. The open-source code of MASAE is available at https://github.com/AdrienLin1/MASAE.

Fetal heart rate (FHR) signals are widely used for fetal health assessment in clinical settings, making them popular in artificial intelligence-based algorithms for fetal health diagnosis. However, a major challenge for such algorithms is the need for a large amount of labeled and category-balanced clinical data to train the models. Like other medical data, FHR faces severe class imbalance in pathological data. Therefore, this paper proposes a minority sample generation method to generate high-quality pathological FHR signals to improve downstream classification task performance. We propose a long time series progressive growing generative adversarial network, TSP-GAN, which dynamically increases the network during training to achieve a transition from coarse-grained to fine-grained time features, thus generating long-time series with rich detailed information. The loss function of this network introduces L2 regularization on the basis of Wasserstein distance and gradient penalty terms to generate high-fidelity signals while avoiding mode collapse. On the one hand, visual and quantitative comparison experiments are designed and the results show that signals of different lengths generated by our network all obtained superior performance. On the other hand, downstream classification tasks are designed and the results indicate that the augmented category-balanced dataset improved by 10% in accuracy compared to the original unbalanced dataset. Therefore, TSP-GAN developed in this paper has practical application value in addressing the problem of sample imbalance in time series.

Axial piston pump fault diagnosis plays a critical role in industrial application field. However, the existing methods face tremendous difficulties in disposing of multi-sensor data-driven and varied pressure pulsation issues. This makes it impossible to select effective diagnostic evidence from multi-sensor data, and the reserved diagnosis model trained under known pressure pulsation fails to adapt for new operation condition. Dedicated to these problems, this paper proposes the manifold transfer (MT) and ensemble filter strategy (EFS) for pump fault diagnosis. In this work, MT is constructed for dimension reduction and feature transformation based on curvilinear component analysis (CCA). It is capable of nonlinear manifold learning to address the issue of varied pressure pulsation. Then, an ensemble filter strategy with an information filtrate function is designed to improve the fault diagnosis performance. The effectiveness of the proposed method is validated by a fault experiment on axial piston pump. The experimental results demonstrate that compared with other existing methods, the proposed method is competitive in terms of diagnostic accuracy and efficiency. Highlights. The manifold transfer is proposed to solve the varied pressure pulsation issue. An ensemble filter strategy is devised to achieve accurate and efficient fault diagnosis without manual intervention. Axial piston pump fault simulation experiments are conducted to validate the effectiveness of proposed method.

Medical imaging is crucial for clinical decisions, but manual report writing is time-consuming and error-prone. Automated generation helps professionals avoid errors and save time. Transformer models excel in capturing long-term dependencies, but face challenges in detailed image representation and modeling lengthy descriptions in medical reports. Addressing these challenges, we propose two modules. First, a transformable and context-aware attentive visual extractor (TCAVE) enhances image features by selecting informative spatial-semantic details across scales, improving report quality. TCAVE involves two novel networks in the ResNet 101 architecture, i.e., (i) incorporating a transformable network into ResNet 101’s intermediate layer for spatially invariant features and (ii) integrating a context-aware network into the final ResNet 101 layer for rich multi-scale contextual features. The transformer encoder encodes fine-grained radiology image details using these representations. Second, a selective memory module with relational gating SMMRG is integrated into the transformer decoder to efficiently model long-term dependencies between input-output sequences while maintaining contextual memory. Our model outperforms existing works on the IU X-Ray and MIMIC-CXR datasets.

Action Quality Assessment (AQA) for diving demands integrated features of human movements and environmental interactions, yet existing works face challenges of insufficient skeleton data and underutilized multi-modal complementarity. To address these, this study first proposes the FineDiving-Pose dataset. By transforming skeletons into stacked pseudo-heatmaps, we achieve unified spatial-temporal modeling with RGB data, avoiding reliance on specialized graph neural networks. We then propose a staged pre-training strategy (decoupling action recognition and AQA) and optimize temporal sampling (replacing redundant two-step regular sampling), enhancing baseline TSA performance while reducing input length. Additionally, we modify the Non-local operator into a dual-input attention fusion block for early RGB-Pose interaction. Qualitative analysis via spatiotemporal attention heatmaps (Dive 405B) shows Pose excels at interference-free human tracking, while RGB captures environmental cues (e.g., water splashes). Quantitative experiments on FineDiving-Pose demonstrate the method outperforms single-modal/baseline models in (varvec{rho }) and (varvec{R})-(varvec{l}_{varvec{2}}), with Pose’s sparse pseudo-heatmaps ensuring faster inference than RGB. Limitations and future directions (e.g., feature alignment for pre-trained weight reuse, generalization to other AQA domains) are also discussed, providing a lightweight, efficient framework for multi-modal AQA.

Clustering is a significant and complex endeavor in machine learning. Symmetric non-negative matrix factorization (SNMF) has attracted considerable interest for its capacity to inherently capture the clustering structure prevalent in graph representation. However, existing SNMF algorithms suffer from issues such as the absence of learning rate and nonlinear learning strategies. To address these issues, this paper proposes a deep symmetric non-negative matrix factorization (DSNMF) representation algorithm for clustering. This algorithm organically integrates the nonlinear strategies of deep learning with the optimization method of SNMF. Specifically, the algorithm focuses on matrix elements and constructs a DSNMF deep network based on non-negative nonlinear constraints and neural network principle. Based on this network, the objective function is minimized. Finally, we evaluated the method on twelve publicly available datasets, including facial recognition images, object images, news text, and biological data. DSNMF achieved favorable clustering performance across these datasets.

We introduce FMC-Net, a facial expression recognition (FER) framework that leverages the hierarchical relationship between discrete facial muscle movements, known as Action Units (AUs), and Facial Expressions (FEs) by integrating two complementary constraint layers. This framework couples data-driven learning with psychology-grounded structure. First, a training-time correlation constraint aligns the two tasks within a multi-task network by softly regularizing a target statistical relationship. This can improve sample efficiency and generalization, particularly under limited or biased data. Second, an inference-time fuzzy rule layer maps the networks probabilistic AU predictions to FEs using compact, human-editable from psychological research, yielding transparent, per-decision attributions. An ensemble then combines the model and rule-based pathways and exposes a disagreement-based risk score for human-in-the-loop triage. This two-layer constraint integration addresses the limitations of single-mechanism approaches: training-time constraints shape the learned representations but lack case-wise transparency, while inference-time rules explain decisions but cannot improve the underlying features. Experiments across diverse datasets, including in-the-wild video and cross-dataset evaluation, validate our approach. Constraint-guided training consistently produces models that outperform competitive baselines, while the rule-based pathway can provide transparency and actionable risk signals towards reliable deployment. The proposed methodology is also generalizable to other machine learning tasks with interdependent outputs.

The development of a safe and accurate biometric system presents a global challenge. Various biometric modalities have been discussed in the literature, the most popular being fingerprint, iris, and voice recognition. While these traditional methods can theoretically achieve a zero error rate, they are not secure. In contrast, a biometric retina identification system offers the highest level of security. This paper introduces an innovative biometric identification method leveraging deep learning and retinal data, explicitly employing the DenseNet architecture. To enhance the model performance, we employ the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), which dynamically adjusts the learning rate, dropout rate, and the number of neurons in the Dense layer. Our research uses several publicly available databases, including the Retinal Identification Database (RIDB), Automated Retinal Image Analysis (ARIA), Structured Analysis of the Retina (STARE), Digital Retinal Images for Vessel Extraction (DRIVE), and Visual Acuity and Retinal Image Analysis (VARIA). Additionally, we have created a new retinal fundus image dataset using an EIDON non-mydriatic retinal camera, providing high-resolution and accurate imaging across multiple modalities. Obtained results present high performance, achieving accuracy rates of 100%, 100%, 99.9%, 100%, 100%, and 100% across the RIDB, ARIA, STARE, DRIVE, VARIA, and our newly collected datasets. These findings highlight the vital role that DenseNet plays in strengthening the security and reliability of personal identification systems. Furthermore, our research emphasizes the importance of the DenseNet architecture in improving the performance and dependability of biometric retina identification systems, thereby enabling secure identification for various applications.

Text classification, which involves the automatic assignment of texts to specific categories, has a broad range of applications in the real world. However, many existing approaches rely on pre-trained language models that, while efficient in learning global linguistic patterns, may struggle with mapping abstract labels to textual data. Additionally, concerns have been raised regarding the robustness of these models and the lack of transparency in their decision-making processes. To address these issues, this paper introduces a novel two-stage learning model for text categorization, which is based on granular-ball representation (TSM-GBR). Initially, texts are transformed into embedding vectors, followed by the generation of granular-balls based on these vectors. Subsequently, a hierarchical strategy based on three-way decision is devised to compute the semantic information of labels. The concept of text confidence is introduced to address samples that the granular-ball model is unable to classify effectively. In the subsequent stage, the semantic representation of word embeddings is refined based on the actual semantics of the labels, with further classification of texts that exhibit low confidence. Considering the limitations of deep learning models in processing semantic information through a single granularity, a dual channel pooling model is designed, which utilizes the max-pooling and the mean-pooling to extract multi-granularity information from the text. Compared with the baseline methods, the proposed model exhibits competitive performance in terms of accuracy and F1-score across various datasets. Extensive comparative experiments confirm that the comprehensive integration of label information significantly enhances text classification. The source codes are available at https://gitee.com/TomisHy/tsm-gbr/tree/master/.

Efficient data communication represents a challenge in wireless data transmission and reception systems, especially in scenarios with intersymbol interference and noise in the communication channel, which degrade the quality of the received information. Two methods are commonly employed to address interference and noise problems: ((varvec{i})) equalization methods to compensate for signal distortions and ((varvec{ii})) error detection methods to identify corrupted data packets. However, these methods have limitations: neural network-based equalization methods categorize all patterns, even in uncertain scenarios, compromising performance, and error detection algorithms are prone to failure, especially in high-noise scenarios. In this paper, we propose a neural network with reject option that simultaneously provides benefits for equalization and error detection. Our approach offers advantages over conventional methods by classifying signals on the basis of their confidence levels. The reject option technique introduced in this paper enhances neural network performance by avoiding classifications with a high risk of error. First, we analyze our proposed neural network using three conventional neural networks for channel equalization. The results indicate that our approach improves performance metrics. After that, we analyze our neural network with a state-of-the-art algorithm by examining bit error rate curves for various communication channels. Finally, we present the results of our method, which are compared with established error detection methods and real data from hardware simulations. Our proposed method can be applied to detect errors without additional data overhead and outperforms other techniques in high-noise scenarios.