International Journal of Computational Intelligence Studies (6 papers in press)

Regular Issues

• Entity-driven attention: lightweight adaptation through frequency-guided scaling  
  by Shifa Pan, Mianzhu Peng, Shiyan Zheng, Chengquan Zhu, Jinmei Wu, Jinfa Wei 
  Abstract: Transformers allocate attention uniformly even when local information density varies substantially across tokens. Existing adaptive attention approaches often introduce additional parameters or complex routing, which is undesirable for resource-constrained deployment. We propose a zero-parameter entity frequency-guided scaling method that adjusts attention logits using normalised token frequency statistics within the context window. Across five repeated shakespeare char runs, the method reduces validation loss from 1.4740 to 1.4631 (0.74% relative improvement; absolute gain 0.0109, 95% CI: +-0.00076) with unchanged parameter count, while additional experiments on text8 and enwik8 show small and dataset-dependent effects. These results indicate that simple statistical adaptation can provide consistent gains on character-level data, and also clarify where this strategy does not generalise.
  Keywords: transformers; attention scaling; information density; parameter-efficient adaptation; character-level language modelling.
  DOI: 10.1504/IJCISTUDIES.2026.10078512
   
  
• Vibration signal based on deep learning in wind turbine state monitoring  
  by Yuanyuan Li, Yingxing Wang, Huan Zhou 
  Abstract: Traditional fault identification methods require a lot of time and effort, and may also miss some important fault features. Moreover, when using complex fault modes and nonlinear data to identify the condition of wind turbines, the accuracy of traditional fault identification methods is relatively low. Therefore, this paper develops a method that uses deep learning to automatically extract features from vibration signals of wind turbines and provide state monitoring and fault diagnosis. This paper first pre-processes the vibration signal by denoising, anomaly detection, normalisation, etc. and then extracts features from the vibration signal effectively improving the quality of the data and providing a better training model. The results indicate that by providing an average recognition accuracy of over 95% on multiple fault types, the overall performance of the model has been improved, which is at least 10% higher than traditional methods.
  Keywords: wind turbine; condition monitoring; deep learning; vibration signal analysis; fault detection.
  DOI: 10.1504/IJCISTUDIES.2026.10079445
   
  
• Student behaviour detection in smart English classroom based on multimodal feature fusion  
  by Liqin Xu, Xiwen Jia, Wei Wang, Wenyan Li, Bin Zhang 
  Abstract: To improve the accuracy and robustness of student behaviour recognition in English teaching in a smart classroom environment, and to promote precise evaluation and intelligent intervention in the English teaching process, this study proposes a student behaviour detection method based on multimodal feature fusion. By collecting and integrating visual, auditory, and interactive multi-source data, a multimodal feature extraction and fusion framework is constructed. A 3D residual network extracts video spatiotemporal features, and temporal and statistical features are extracted from interaction logs. An LSTM network models the fused temporal features to capture the dynamic evolution of student behaviour, achieving automatic recognition and classification of various learning behaviours and abnormal behaviours such as focused listening, group discussions, and active responses. Experimental results show that this method achieves a detection accuracy of over 90%, a detection error rate of less than 10%, and a recall rate above 0.9.
  Keywords: multimodal feature fusion; smart classroom; English teaching; Student behaviour detection; LSTM.
  DOI: 10.1504/IJCISTUDIES.2026.10079661
   
  
• Automatic annotation of pronunciation errors in Chinese teaching based on CNN recognition model of pronunciation features  
  by Mingkun Zhu, Jinhua Xiong 
  Abstract: In teaching Chinese as a second language, timely diagnosis and correction of pronunciation errors are key to improving teaching efficiency and learning outcomes. Therefore, a method for automatic annotation of pronunciation errors in Chinese language teaching based on a convolutional neural network recognition model using pronunciation features is proposed. Extract multi-dimensional acoustic features such as Mel frequency cepstral coefficients, fundamental frequency, and sound intensity from speech signals, and construct a convolutional neural network bidirectional long short-term memory network multitask learning model to achieve phoneme classification and pronunciation feature recognition. By combining Bayesian decision theory and adaptive threshold determination mechanism, the types and severity levels of phonemes with pronunciation errors are determined, and the final output is a clear and interpretable pronunciation error annotation result. The experimental results show that this method achieves high accuracy in detecting errors, and its annotation results have good consistency with expert annotations.
  Keywords: pronunciation error detection; CNN; pronunciation characteristics; automatic annotation; Chinese language teaching.
  DOI: 10.1504/IJCISTUDIES.2026.10079663
   
  
• Beyond sequential prediction: learning financial market dynamics in volatile and non-stationary environments through sentiment-conditioned generative modelling  
  by Alexis Lazanas, Spyridon Karpouzis 
  Abstract: The problem of time-series forecasting in non-stationary and complex environments is a challenging task in machine learning, especially with heterogeneous numerical and textual data present. Traditional statistical models like autoregressive integrated moving average (ARIMA) are based on the assumptions of linearity and stationarity, whereas recurrent neural networks like long short-term memory (LSTM) models do not necessarily represent distributional properties in highly volatile settings. This paper proposes a hybrid model that combines generative adversarial networks (GANs) with natural language processing (NLP)-based sentiment analysis to enable sentiment-conditioned time-series prediction. The model integrates adversarial learning on numerical sequences with contextual sentiment representations derived from unstructured text, enabling them to be jointly modelled to capture temporal dynamics and exogenous information. These results demonstrate the promise of hybrid generative and language-aware methods to enhance prediction robustness in non-stationary environments.
  Keywords: computational intelligence; generative models; sentiment analysis; financial time-series forecasting; volatile and non-stationary environments.
  DOI: 10.1504/IJCISTUDIES.2026.10079804
   
  
• Application of improved DBSCAN and visualisation techniques in clustering typical vessel trajectories in port waters  
  by Rong Ge 
  Abstract: To meet demands for clustering and visualisation of ship trajectories in port waters, this paper proposes an improved density-based spatial clustering algorithm with noise filtering for typical trajectory extraction. Built on traditional methods, it defines core distance and fusion distance to improve clustering accuracy in complex port environments. Core distance adaptively recognises trajectory density and reduces parameter dependency. Fusion distance integrates spatial positions and total trajectory length to measure similarity, distinguish different sailing behaviours and greatly boost clustering performance. Combined with visualisation, it clearly displays main vessel routes. Experiments show its clustering accuracy reaches 0.947 and F1 score 0.928. It takes merely 0.91 seconds to process 150,000 data samples, surpassing other algorithms. Visual results clearly reflect path differences of various navigation patterns and typical route features. This algorithm provides a feasible solution for port trajectory clustering and offers solid technical support and decision references for smart port intelligent management.
  Keywords: DBSCAN; visualisation techniques; port waters; vessel trajectories; automatic identification system; AIS.
  DOI: 10.1504/IJCISTUDIES.2026.10079977