HWHeshan Wang

Heshan Wang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Heshan Wang has a total of 54 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • KT-RC: Kernel Time-Delayed Reservoir Computing for Time Series Prediction

    by Heshan Wang, Mengmeng Chen, Kunjie Yu, Jing Liang, Zhaomin Lv, Zhong Zhang

    Reservoir computing (RC) is an efficient recurrent neural network (RNN) method. However, the performance and prediction results of traditional RCs are susceptible to several factors, such as their network structure, parameter setting, and selection of input features. In this study, we employ a kernel time-delayed RC (KT-RC) method for time series prediction. The KT-RC transforms input vectors linearly to obtain a high-dimensional set of time-delayed linear eigenvectors, which are then transformed by various kernel functions to represent the nonlinear characteristics of the input signal. Finally, the Bayesian optimization algorithm adjusts the few remaining weights and kernel parameters to minimize the manual adjustment process. The advantages of KT-RC can be summarized as follows: 1) KT-RC solves the problems of uncertainty in weight matrices and difficulty in large-scale parameter selection in the input and hidden layers of RCs. 2) The KT module can avoid massive reservoir hyperparameters and effectively reduce the hidden layer size of the traditional RC. 3) The proposed KT-RC shows good performance, strong stability, and robustness in several synthetic and real-world datasets for one-step-ahead and multistep-ahead time series prediction. The simulation results confirm that KT-RC not only outperforms some gate-structured RNNs, kernel vector regression models, and recently proposed prediction models but also requires fewer parameters to be initialized and can reduce the hidden layer size of the traditional RCs. The source code is available at https://github.com/whs7713578/RC.

  • Open Access

    Group Method of Data Handling Spiking Neural Network with Variational Mode Decomposition for Pattern Classification

    by Heshan Wang, Liangkun Guo, Bowen Sun, Jiangtao Wang, Tianlei Ma, Jinzhu Peng, Sheng Shang

    As the data features acquired nowadays are becoming increasingly complex and distinctive, traditional single-layer spiking neural networks (SNNs) are increasingly struggling to train such data. Deep SNN (DSNN) can effectively provide biological representations and effective architectures for the above-mentioned data. The weights of traditional deep artificial neural networks (DANN) can be transferred into those of equivalent DSNN via conversion method. However, the shortcomings such as the non-differentiability of spiking information, poor interpretability, and excessive randomness make it challenging to train DSNN using backpropagation-based techniques. To improve the performance of DSNN, enhance its feature learning capability, and reduce the complexity of parameter design, this paper proposes a hybrid DSNN structure integrated with group method of data handling (GMDH) network and an SNN classifier for pattern classification tasks. In the hybrid GMDH-SNN model, complex categorical data are first decomposed into several simplified modes via Variational mode decomposition (VMD). Subsequently, the multi-layer GMDH module can effectively extract discriminative features from the decomposed input data through self-organizing inductive learning. This design not only improves the interpretability and reduces the structural complexity of DSNN, but also achieves competitive recognition accuracy comparable to DANN. Extensive classification benchmarks are used to validate the effectiveness of the GMDH-SNN model. Nonparametric statistical test, feature visualization results and experimental outcomes reveal that the GMDH-SNN with VMD not only outperforms some traditional classifiers, recently proposed ANN-based classifiers, DSNN classifiers, and GMDH-based classifier, but also reduces the complexity of DSNN.

  • GVA-BLS: Gaussian Vector Autoregression Broad Learning System Based on Randomly Distributed Embedding for Multistep-Ahead Prediction

    by Mingyuan Yu, Bowen Sun, Heshan Wang, Liangkun Guo, Jinzhu Peng

    As a model proficient in time series prediction, the recurrent broad learning system (RBLS) combines the advantages of traditional BLS and recurrent neural networks (RNNs). This combination enhances the learning of transient data and ensures computational efficiency. However, RBLS still suffers from the shortcomings of traditional RNNs. These include excessive connection weights, an unexplainable model structure, and low computational efficiency. To improve recurrent neurons and structure in RBLS, this article proposes a Gaussian vector autoregression BLS (GVA-BLS) model based on nonlinear vector autoregression and RBLS theory. GVA-BLS is applied to multistep-ahead time series prediction. Compared to traditional RBLS, GVA-BLS has three advantages: 1) GVA-BLS addresses the issue of excessive recurrent weights in RBLS enhancement nodes and the difficulty in random initialization of hyperparameters; 2) the GVA nodes use a Gaussian kernel function to perform a nonlinear transformation on the output of feature nodes, which makes the GVA-BLS more efficient; and 3) compared to the traditional enhancement nodes of RBLS, GVA nodes can map the output of feature nodes into a higher dimensional space, making GVA-BLS more suitable for short-term, high-dimensional, multistep-ahead prediction tasks. Experimental results demonstrate that GVA-BLS outperforms several recently proposed gate-structured RNNs and BLS-based models in terms of multistep-ahead prediction performance. It also reduces the computational overhead of RBLS and makes it more efficient and intuitive.

  • Historical Information-Assisted Dynamic Response Integration and Adaptive Niche Methods for Dynamic Multimodal Optimization

    by Kunjie Yu, Xuyang Zhang, Dezheng Zhang, Jing Liang, Yumeng Li, Heshan Wang, Ke Chen, Caitong Yue

    Dynamic multimodal optimization problems (DMMOPs) represent the multimodal optimization problems that the optimal solution changes over time. Due to the wide application of DMMOPs in reality, some related algorithms have been proposed in recent years. Most existing algorithms employ a single dynamic response mechanism and embed it in existing multi-modal evolutionary algorithms. However, these algorithms often perform limited when environmental change involves multiple types, and they fail to consider utilizing historical information to assist static multimodal optimizers. To solve these issues, this paper proposes historical information-assisted dynamic response integration and adaptive niche methods (HIA-DRI-AN) for dynamic multi-modal optimization. In HIA-DRI-AN, an dynamic response integration method with adaptive adjustment mechanism is proposed for generating the initial population when the change happens. This method considers the change types of DMMOPs, and integrates targeted dynamic response mechanisms to respond to the different change types. Also, this method can adaptively self-adjust to balance the convergence and diversity of the initial population depending on the integrated response mechanism’s performance in historical environments. Furthermore, a niching adaptive division strategy is proposed to enhance the performance of the static optimizer. The strategy dynamically divides niches based on the integrated response mechanism’s performance and the current evolutionary stage, which can adjust the preference for diversity and convergence during evolution. The comprehensive experimental results on 24 test functions show that HIADRI-AN is superior compared to some state-of-the-art dynamic multimodal algorithms.


  • DMEF-Net: Lightweight Infrared Dim Small Target Detection Network for Limited Samples

    by Tianlei Ma, Zhen Yang, Yi- Fan, Song, Jing Liang, Heshan Wang

    Intractable sparse feature extraction, over-weight model size, and limited training samples are currently bewildering in infrared dim small target detection, which are not adequately addressed by current state-of-the-art (SOTA) methods. Here, to synchronously address these issues, a dense multi-level feature extraction and fusion network (DMEF-net) is designed, mainly consisting of two modules: target context and Gaussian saliency feature extraction module (TCGS) and multi-level dense feature fusion structure (MLDF). Inspired by the physical thermal diffusion model and the human visual mechanism for small-scale targets, Gaussian salience features and local context features are introduced into the target feature expression through the designed TCGS module to solve the sparse feature extraction problem. Then, to solve the over-weight model size problem, a novel MLDF module is designed to incorporate the feature reuse mechanism into our model, thereby significantly reducing the number of trainable parameters. Finally, in the training procedure, an efficient saliency labeling strategy is proposed to jointly supervise the model training in both Euclidean and Gaussian spaces, ultimately enhancing the model’s ability to explore the target features and alleviate performance deterioration when the training samples are limited. Extensive experiments on several large-scale open datasets, including TDSATUA and MTDUCB, prove that the proposed DMEF-net outperforms other SOTA methods by 8% in accuracy and has 4800% less model size.


  • Open Access

    Bearing-Fault Diagnosis with Signal-to-RGB Image Mapping and Multichannel Multiscale Convolutional Neural Network

    by Ming Xu, Jinfeng Gao, Zhong Zhang, Heshan Wang

    Deep learning bearing-fault diagnosis has shown strong vitality in recent years. In industrial practice, the running state of bearings is monitored by collecting data from multiple sensors, for instance, the drive end, the fan end, and the base. Given the complexity of the operating conditions and the limited number of bearing-fault samples, obtaining complementary fault features using the traditional fault-diagnosis method, which uses statistical characteristic in time or frequency, is difficult and relies heavily on prior knowledge. In addition, intelligent bearing-fault diagnosis based on a convolutional neural network (CNN) has several deficiencies, such as single-scale fixed convolutional kernels, excessive dependence on experts’ experience, and a limited capacity for learning a small training dataset. Considering these drawbacks, a novel intelligent bearing-fault-diagnosis method based on signal-to-RGB image mapping (STRIM) and multichannel multiscale CNN (MCMS-CNN) is proposed. First, the signals from three different sensors are converted into RGB images by the STRIM method to achieve feature fusion. To extract RGB image features effectively, the proposed MCMS-CNN is established, which can automatically learn complementary and abundant features at different scales. By increasing the width and decreasing the depth of the network, the overfitting caused by the complex network for a small dataset is eliminated, and the fault classification capability is guaranteed simultaneously. The performance of the method is verified through the Case Western Reserve University’s (CWRU) bearing dataset. Compared with different DL approaches, the proposed approach can effectively realize fault diagnosis and substantially outperform other methods.


  • An Incremental Learning Model for Mobile Robot: From Short-Term Memory to Long-Term Memory

    by Dongshu Wang, Kai Yang, Lei Liu, Heshan Wang

    During the environmental cognition, how to realize the efficient incremental learning of a mobile robot is a great challenge. The existing methods suffer from the low efficiency. This article proposes a novel methodology to address it. A memory model inspired by the human brains is constructed to realize the transmission of short-term memory (STM) to long-term memory (LTM) in the offline states, thus realize the efficient incremental learning of the mobile robot. Concretely, during the online process, when the sensory information is input to the developmental network (DN), similarity between the new input information and the knowledge memorized in the DN is calculated first. If the similarity is larger than the threshold, this input is transmitted to motor layer and determine the optimal decision. Otherwise, the sensory input will be temporarily stored in a neuron in the STM by an evaluation function. During the offline states, a self-triggering mechanism is designed to trigger the DN to work again. Then, the lateral excitation of the internal neurons is designed to fire more neurons to memorize the knowledge transferred from the STM. Then, the synaptic weights of the new fire neurons are updated, and the STM becomes the LTM, thus realizing the incremental learning. In the following task, if the robot encounters a similar scenario, it can make a quick decision, based on the knowledge learned during the offline states. This continuous learning pattern reduces the training samples used in the network, hence enhancing its learning efficiency. Most importantly, this methodology makes the robot continuously improve its intelligence through the incremental learning, even in off-line states. Extensive simulation and experiment results of the mobile robot navigation demonstrate its potential.