Search

Underwater image processing is an intelligence research field that has great potential to help developers better explore the underwater environment. Underwater image processing has been used in a wide variety of fields, such as underwater microscopic detection, terrain scanning, mine detection, telecommunication cables, and autonomous underwater vehicles. However, underwater imagery suffers from strong absorption, scattering, color distortion, and noise from the artificial light sources, causing image blur, haziness, and a bluish or greenish tone. Therefore, the enhancement of underwater imagery can be divided into two methods: 1) underwater image dehazing and 2) underwater image color restoration. This paper presents the reason for underwater image degradation, surveys the state-of-the-art intelligence algorithms like deep learning methods in underwater image dehazing and restoration, demonstrates the performance of underwater image dehazing and color restoration with different methods, introduces an underwater image color evaluation metric, and provides an overview of the major underwater image applications. Finally, we summarize the application of underwater image processing.

High-dimensional and large-scale time series processing has aroused considerable research interests during decades. It is difficult for traditional methods to reveal the evolution state in dynamical systems and discover the relationship among variables automatically. In this paper, we propose a unified framework for nonuniform embedding, dynamical system revealing, and time series prediction, termed as Structured Manifold Broad Learning System (SM-BLS). The structured manifold learning is introduced for nonuniform embedding and unsupervised manifold learning simultaneously. Graph embedding and feature selection are both considered to depict the intrinsic structure connections between chaotic time series and its low-dimensional manifold. Compared with traditional methods, the proposed framework could discover potential deterministic evolution information of dynamical systems and make the modeling more interpretable. It provides us a homogeneous way to recover the chaotic attractor from multivariate and heterogeneous time series. Simulation analysis and results show that SM-BLS has advantages in dynamic discovery and feature extraction of large-scale chaotic time series prediction.

State space reconstruction is the foundation of chaotic system modeling. Selection of reconstructed variables is essential to the analysis and prediction of multivariate chaotic time series. As most existing state space reconstruction theorems deal with univariate time series, we have presented a novel nonuniform state space reconstruction method using information criterion for multivariate chaotic time series. We derived a new criterion based on low dimensional approximation of joint mutual information for time delay selection, which can be solved efficiently through the use of an intelligent optimization algorithm with low computation complexity. The embedding dimension is determined by conditional entropy, after which the reconstructed variables have relatively strong independence and low redundancy. The scheme, which integrates nonuniform embedding and feature selection, results in better reconstructions for multivariate chaotic systems. Moreover, the proposed nonuniform state space reconstruction method shows good performance in forecasting benchmark and actual multivariate chaotic time series.

Time series prediction is an active area and attracts the attention of researchers from lots of fields. With the expansion of the application of time series forecasting in the real world, many traditional methods that predicted time series with a few data are no longer applicable. Motivated by the demand of accurate prediction, we come up with the piecewise time series prediction model combining stacked long short-term memory network with genetic algorithm. Stacked long short-term memory network (LSTM) is an advanced sequence learning technology. We divide the time series into several segments and then use the stacked LSTM trained on the previous segment to predict the next segment to realize online prediction. The parameters of the stacked LSTM for each segment prediction are optimized by improved genetic algorithm. For purpose of demonstrating the feasibility and effectivity of the proposed model, we carry out simulation experiments on basis of a hybrid benchmark chaotic time series and a real-world dataset of the hourly ozone concentration. Experimental results demonstrate that the proposed model can automatically select the proper structure according to the data, which avoids the structure being too redundant or too simple. Furthermore, it can improve the prediction accuracy of time series and has good practicability.

In order to solve the problem of the speech recognition, this paper proposes a novel framework to process speech signals by using the deep fully convolutional neural network (DFCNN) as the acoustic model and the Transformer as the linguistic model. DFCNN can learn more about speech signals through a deep network structure, while Transformer's self-attention mechanism makes it an efficient linguistic model. In order to better extract the characteristic information of speech signal, we transform the speech signal into a two-dimensional language spectrum by frame division and Fourier transform. In this way, the language spectrum can be processed by using convolutional neural network, which has promising ability to process two-dimensional data. The output of the acoustic model is passed into the linguistic model, and a complete sentence is output. At the same time, considering that there are many hyper parameters in the model, in order to adjust these hyper parameters efficiently, we employ the Bayesian optimization algorithm to optimize the hyper parameters of the acoustic model and the linguistic model, so as to further improve the accuracy of speech recognition. Experimental results on speech recognition library THCHS30 show that compared with other models, the optimized model has higher recognition rate for audio and stronger generalization ability.

Visual saliency (VS) is an important mechanism for defining which areas of an image will attract more attention of the HVS. Thus, VS can be employed to weight the just noticeable difference (JND) with different attention levels. Some VS-based JND profiles have been proposed in the DCT domain, which used the bottom-up features, such as luminance and texture only. Recently, the research about saliency detection has shown that a better saliency model considering both bottom-up features and top-down features will lead to a significant improvement for the overall saliency detection performance. In this paper, we propose a novel two-layer VS-induced JND profile, which is composed of the bottom-up features and the top-down feature extracted from DCT blocks in the transformed domain. In this model, the luminance and texture features are adapted to calculate the bottom-up features maps, while the top-down feature of focus is used to guide the generation of final salient regions since the camerapersons are used to facilitate the attention regions in focus. The proposed two-layer saliency-induced JND model is further applied to modulate the quantization step in the watermarking framework, which can make full use of its individual merits to achieve a better tradeoff between fidelity and robustness. The experimental results show that the proposed scheme has superior performance than the previous watermarking schemes.

Motivated by the idea of 'decomposition and ensemble', this paper proposes a novel method based on the wavelet-denoising algorithm and multiple echo state networks to improve the prediction accuracy of noisy multivariate time series. The noisy time series is first denoised by a wavelet soft thresholding algorithm and decomposed into a set of well-behaved constitutive series. Each constitutive series is then predicted by a separate echo state network with proper parameters that match the specified dynamics. Finally, the overall prediction is achieved by a linear combination of the constitutive series. For each constitutive series, we use the correlation integral method to select the phase-reconstruction parameters and to construct the appropriate input. Two sets of multivariate time series are investigated using the proposed model and some other related work. The simulation results demonstrate the effectiveness of the proposed method.

With the applications of Internet of Things (IoT) for smart cities, the real-time performance for a large number of network packets is facing serious challenge. Thus, how to improve the emergency response has become a critical issue. However, traditional packet scheduling algorithms cannot meet the requirements of the large-scale IoT system for smart cities. To address this shortcoming, this paper proposes EARS, an efficient data-emergency-aware packet scheduling scheme for smart cities. EARS describes the packet emergency information with the packet priority and deadline. Each source node informs the destination node of the packet emergency information before sending the packets. The destination node determines the packet scheduling sequence and processing sequence according to emergency information. Moreover, this paper compares EARS with a first-come, first-served, multilevel queue algorithm and a dynamic multilevel priority packet scheduling algorithm. Simulation results show that EARS outperforms these previous scheduling algorithms in terms of packet loss rate, average packet waiting time, and average packet end-to-end delay.

In this paper, a novel method (Hybrid–ELM–PLSR) is proposed based on hybrid variable selection algorithm and improved extreme learning machine (ELM) for multivariate chaotic time series prediction. The hybrid variable selection algorithm combines the advantages of filter and wrapper, effectively balancing the calculation speed and prediction accuracy. Moreover, for ELM, multicollinearity, which can result in ill-condition, is always existent among the hidden layer output matrix. And the optimal number of hidden nodes is also difficult to be determined. Therefore,in order to overcome these problems, an improved ELM (ELM–PLSR) is proposed based on partial least square regression (PLSR). It can effectively enhance the stability performance and prediction performance of ELM. Hybrid–ELM–PLSR can be divided into three stages. At first, filter is used to rearrange the input variables through the correlations with desired variables. Then wrapper is used to select the optimal variable subset through evaluating the prediction performance of different subsets. Finally, ELM–PLSR is used to build the prediction model. The simulation experiment results based on San Francisco river runoff dataset demonstrate that the proposed method is effective for multivariate chaotic time series. And the prediction accuracy and reliability are higher than other methods.

Accurate prediction of solar irradiance is beneficial in reducing energy waste associated with photovoltaic power plants, preventing system damage caused by the severe fluctuation of solar irradiance, and stationarizing the power output integration between different power grids. Considering the randomness and multiple dimension of weather data, a hybrid deep learning model that combines a gated recurrent unit (GRU) neural network and an attention mechanism is proposed forecasting the solar irradiance changes in four different seasons. In the first step, the Inception neural network and ResNet are designed to extract features from the original dataset. Secondly, the extracted features are inputted into the recurrent neural network (RNN) network for model training. Experimental results show that the proposed hybrid deep learning model accurately predicts solar irradiance changes in a short-term manner. In addition, the forecasting performance of the model is better than traditional deep learning models (such as long short term memory and GRU).

Just noticeable distortion (JND) is widely employed to describe the perception redundancy in the quantization-based watermarking framework. However, the existing JND models are generally constructed to treat every region of the photograph with an equal focus level, whereas the defocus effect has never been considered. In this paper, the defocus feature, which can portray the aesthetic emphasis in the photograph, is provided to improve the perceptual JND model. Firstly, two indicators which consider the block energy in the defocus measurement (DM) are proposed. Then, the defocus feature map (DFM) is obtained by integrating the influence of the circumambient blocks, and it is applied to the proposed JND contrast masking (CM) processing. In this way, a new blind photograph watermarking method, with emphasis on defocus-JND estimation combined with the proposed CM, is presented. Simulations show that the proposed JND is more suitable for watermarking framework than some exiting JND models, and the proposed watermarking scheme with the improved defocus-based JND model has superior robustness compared with some watermarking schemes.

Ground surface settlement forecasting in the process of tunnel construction is one of the most important techniques towards sustainable city development and preventing serious damages, such as landscape collapse. It is evident that modern artificial intelligence (AI) models, such as artificial neural network, extreme learning machine, and support vector regression, are capable of providing reliable forecasting results for tunnel surface settlement. However, two limitations exist for the current forecasting techniques. First, the data provided by the construction company are usually univariate (i.e., containing only the settlement data). Second, the demand of tunnel surface settlement is immediate after the construction process begins. The number of training data samples is limited. Targeting at the above two limitations, in this study, a novel ensemble machine learning model is proposed to forecast tunnel surface settlement using univariate short period of real-world tunnel settlement data. The proposed Adaboost.RT framework fully utilizes existing data points with three base machine learning models and iteratively updates hyperparameters using current surface point locations. Experimental results show that compared with existing machine learning techniques and algorithms, the proposed ensemble learning method provides a higher prediction accuracy with acceptable computational efficiency.

A better tradeoff between robustness and invisibility will be realized by using the just noticeable (JND) model into the quantization-based watermarking scheme. The JND model is usually used to describe the perception characteristics of human visual systems (HVS). According to the research of cognitive science, HVS can adaptively extract the structure features of an image. However, the existing JND models in the watermarking scheme do not consider the structure features. Therefore, a novel JND model is proposed , which includes three aspects: contrast sensitivity function, luminance adaptation, and contrast masking (CM). In this model, the CM effect is modeled by analyzing the direction features and texture complexity, which meets the human visual perception characteristics and matches well with the spread transform dither modulation (STDM) watermarking framework by employing a new method to measure edge intensity. Compared with the other existing JND models, the proposed JND model based on structural regularity is more efficient and applicable in the STDM watermarking scheme. In terms of the experimental results, the proposed scheme performs better than the other watermarking scheme based on the existing JND models.

Time series prediction is an important branch of mathematical statistics, that uses stochastic process theory and machine learning method to study the data sequence. It is based on the research of structural models, with the purpose of predicting the future developing trend. The method proposed in this paper is based on ARMA model and Lasso regression, ARMA model is an effective model for time series prediction, and Lasso regression is a kind of compression estimation method to refine the ARMA model. First, we build an ARMA model. Second, we use the extend autocorrelation function to do the model specification. Third, we use Lasso to calculate the model parameters. Finally, the method presented in this paper is compared with some other methods on two real-world datasets. Experimental results demonstrate its effectiveness for time series prediction.

The broad learning system (BLS) is an emerging approach for effective and efficient modeling of complex systems. The inputs are transferred and placed in the feature nodes, and then sent into the enhancement nodes for nonlinear transformation. The structure of a BLS can be extended in a wide sense. Incremental learning algorithms are designed for fast learning in broad expansion. Based on the typical BLSs, a novel recurrent BLS (RBLS) is proposed in this paper. The nodes in the enhancement units of the BLS are recurrently connected, for the purpose of capturing the dynamic characteristics of a time series. A sparse autoencoder is used to extract the features from the input instead of the randomly initialized weights. In this way, the RBLS retains the merit of fast computing and fits for processing sequential data. Motivated by the idea of “fine-tuning” in deep learning, the weights in the RBLS can be updated by conjugate gradient methods if the prediction errors are large. We exhibit the merits of our proposed model on several chaotic time series. Experimental results substantiate the effectiveness of the RBLS. For chaotic benchmark datasets, the RBLS achieves very small errors, and for the real-world dataset, the performance is satisfactory.

Improving universality and robustness of the control method is one of the most challenging problems in the field of complex networks (CNs) synchronization. In this paper, a special unilateral coupling finite-time synchronization (UCFTS) method for uncertain CNs is proposed for this challenging problem. Multiple influencing factors are considered, so that the proposed method can be applied to a variety of situations. First, two kinds of drive-response CNs with different sizes are introduced, each of which contains two types of nonidentical nodes and time-varying coupling delay. In addition, the node parameters and topological structure are unknown in drive network. Then, an effective UCFTS control technique is proposed to realize the synchronization of drive-response CNs and identify the unknown parameters and topological structure. Second, the UCFTS of uncertain CNs with four types of nonidentical nodes is further studied. Moreover, both the networks are of unknown parameters, time-varying coupling delay and uncertain topological structure. Through designing corresponding adaptive updating laws, the unknown parameters are estimated successfully and the weight of uncertain topology can be automatically adapted to the appropriate value with the proposed UCFTS. Finally, two experimental examples show the correctness of the proposed scheme. Furthermore, the method is compared with the other three synchronization methods, which shows that our method has a better control performance.

To deal with “curse of dimensionality” caused by high-dimensional data and the underutilization of spatial information in the classification of hyperspectral imagery (HSI), a new spectral-spatial classification method based on probabilistic support vector machine (SVM) and guided filter is proposed. The proposed method utilizes spatial information in two aspects. Firstly, the extended morphological profiles (EMP), differential morphological profiles (DMP) and Gabor textural features of hyperspectral image are extracted, which can preserve the texture, outline and shape of the original image. The spectral and extracted spatial features are joint classified by probabilistic SVM to obtain the initial classification probabilistic maps. Secondly, the guided filter is used to regularize the initial classification probabilistic maps, which can enhance the smoothness of local neighbourhood and preserve the edge information. The proposed algorithm is examined by the Indian Pines and University of Pavia hyperspectral datasets. On the same experimental conditions, the proposed method achieves the highest classification accuracy and the most reasonable classification map, demonstrating the proposed algorithm has obvious advantages in hyperspectral remote image classification.

It has been known that human visual systems (HVSs) can be applied to describe the underlying masking properties for the image processing. In general, HVS can only perceive small changes in a scene when they are greater than the just noticeable distortion (JND) threshold. Recently, the cognitive resources of huma visual attention mechanisms are limited, which can not concentrate on all stimuli. To be specific, only more important stimuli will react from the mechanisms. When it comes to visual attention mechanisms, we need to introduce the visual saliency to model the human perception more accurately. In this paper, we presents a new wavelet-based JND estimation method that takes into account the interrelationship between visual saliency and JND threshold. In the experimental part, we verify it from both subjective and objective aspects. In addition, the experimental results show that extracting the saliency map of the image in the discrete wavelet transform (DWT) domain and then modulating its JND threshold is better than the non-modulated JND effect.

Recently, the scale of observed time series becomes bigger and bigger, and the learning ability of conventional models is limited to depict the characteristics of large-scale time series. We put forward a deep model that uses the sparse autoencoder to extract the characteristics of the time series layer by layer. Sparse autoencoder can provide a simple explanation of the data by means of a reduced dimensionality of parts and extract the hidden architecture in the data. By this method, on the one hand, the dimension of the matrix that describes the problem is reduced, on the other hand, a large number of data can be compressed and summarized. Then we calculate the output weights by the method of pseudo inverse. Experiment results on Lorenz series and PM2.5 concentration in Shanghai dataset demonstrate the effectiveness of the proposed method.

Kernel recursive least squares (KRLS) has become a new research hotspot, attracting researchers’ attention in the field of time series online prediction. As an online algorithm, its computational form is updating incrementally as a new sample achieved. But as the number of data samples increases, its computational complexity will increase. And there is a lack of ability to track time-varying characteristics in the online prediction process. Therefore, we add a sliding window approach to the sparse KRLS and propose a new improved algorithm, which is sparse kernel recursive least squares based on sliding window (SW-SKRLS). This improved algorithm can limit the computational complexity and perform better online prediction in time-varying environments. To prove the validity of the proposed algorithm, we make simulations on the Lorenz time series and Dalian monthly average temperature and rainfall time series. The simulation results show that this algorithm has good prediction accuracy and efficiency.

Extreme learning machines have attracted increasing attention for setting the input weights and hidden biases randomly. However, excessive hidden neurons in ELMs may cause ill-posed problem when Moore-Penrose generalized inverse method is used to train the neural network. It generates large amplitude of output weights, which leads to bad generalization. To solve this problem, we propose an improved multi-objective particle swarm optimization algorithm to optimize the output weights of ELMs. It consists of two objective functions. One is the sum of the errors between predicted values and target values. The other one is the $L_{2}$ norm of output weights. We improve general MOPSO algorithm by adding chaotic mapping to promote the ergodicity and randomness of new solutions. In addition, a new strategy is proposed to select $pbest$ and $gbest$ to deal with the unbalance of multiple objectives. Experimental results on predicting two benchmark chaotic time series sustain the efficiency and effectiveness of our proposed algorithm. It has better generalization performance and smaller output weights than original ELM and other evolutionary ELMs.

Echo state network is a novel kind of recurrent neural networks, with a trainable linear readout layer and a large fixed recurrent connected hidden layer, which can be used to map the rich dynamics of complex real-world data sets. It has been extensively studied in time series prediction. However, there may be an ill-posed problem caused by the number of real-world training samples less than the size of the hidden layer. In this brief, a Laplacian echo state network (LAESN), is proposed to overcome the ill-posed problem and obtain low-dimensional output weights. First, an echo state network is used to map the multivariate time series into a large reservoir. Then, assuming that an unknown underlying manifold is inside the reservoir, we employ the Laplacian eigenmaps to estimate the manifold by constructing an adjacency graph associated with the reservoir states. Finally, the output weights are calculated by the low-dimensional manifold. In addition, some criteria of transient stability, local controllability, and local observability are given. Experimental results based on two real-world data sets substantiate the effectiveness and characteristics of the proposed LAESN model.

In this research, for online modeling and prediction of multivariate time series, we propose a novel approach termed squared root cubature Kalman filter-γ echo state network (SCKF-γESN). First, multivariate time series are modeled by using γ echo state network (γESN). Then, by using squared root cubature Kalman filter (SCKF), we update parameters of γESN and predict future observations online. Furthermore, we add a robust outlier detection algorithm to SCKF to protect SCKF-γESN from divergence caused by outliers. Finally, two numerical examples, by using a multivariate benchmark dataset and a real-world dataset, are conducted to substantiate the effectiveness and characteristics of the proposed SCKF-γESN.

The just noticeable distortion (JND) model plays an important role in measuring the visual visibility for spread transform dither modulation (STDM) watermarking. However, the existing JND model characterizes the suprathreshold distortions with an equal saliency level. Visual saliency (VS) has been widely studied by psychologists and computer scientists during the last decade, where the distortions are more likely to be noticeable to any viewer. With this consideration, we proposed a novel STDM watermarking method for a monochrome image by exploiting a visual saliency-based JND model. In our proposed JND model, a simple VS model is employed as a feature to reflect the importance of a local region and compute the final JND map. Extensive experiments performed on the classic image databases demonstrate that the proposed watermarking scheme works better in terms of the robustness than other related methods.

Echo state network (ESN) is a novel kind of recurrent neural networks, where a reservoir is generated randomly and only readout layer is adaptable. It outperforms conventional recurrent neural networks in the field of multivariate time series prediction. Often ESN works beautifully. But sometimes it works poorly because of ill-posed problem. To solve it, we propose a new model on the basis of ESNs, termed as fast subspace decomposition echo state network (FSDESN). The core of the model is to utilize fast subspace decomposition algorithm for extracting a compact subspace out of a redundant large-scale reservoir matrix in order to remove approximate collinear components, overcome the ill-posed problem, and improve generalization performance. Experimental results on two multivariate benchmark datasets substantiate the effectiveness and characteristics of FSDESN.

The perceptual just noticeable distortion (JND) model has attracted increasing attention in the field of the quantization-based watermarking framework. The JND model can provide a superior tradeoff between robustness and fidelity. However, the conventional JND models are not fit for the quantization-based watermarking, as the image has been altered by watermarking embedding. In this paper, we present an improved spread transform dither modulation (STDM) watermarking scheme, which is based on the image primitive features produced according to JND mechanism. The procedures include the contrast masking effect by utilizing a new measurement of edge strength which represent semantic information. What’s more, the proposed semantic information-based JND model can be theoretically invariant to the changes in the watermark-embedding processing. The newly proposed JND model is very simple but more effective in the STDM watermarking. Experiments results demonstrate that the proposed watermarking scheme can bring about better performance compared with previously proposed perceptual STDM schemes.

Echo state network (ESN) is a new kind of recurrent neural network with a randomly generated reservoir structure and an adaptable linear readout layer. It has been widely employed in the field of time series prediction. However, when high-dimensional reservoirs are utilized to predict multivariate time series, there may be a collinearity problem. In this paper, to overcome the collinearity problem and obtain a sparse solution, we propose a new model-adaptive elastic ESN, in which adaptive elastic net algorithm is used to calculate the unknown weights. It combines the strengths of the quadratic regularization and the adaptively weighted lasso shrinkage. Hence, the proposed model can deal with the collinearity problem and enjoy the oracle property with an unbiased estimation. We exhibit the merits of our model on two benchmark multivariate chaotic datasets and two real-world applications. Experimental results substantiate the effectiveness and characteristics of the proposed model.

In order to effectively utilize the network teaching resources, a teaching resource classification method based on the improved KNN (K-Nearest Neighbor) algorithm was proposed. Taking the text class primary and secondary school teaching resources as the research object, combined with the domain characteristics, the KNN algorithm was improved. By measuring the sample space density, the text of the high-density area was found. Different clipping methods were proposed for both intra-class and inter-class regions. The problem of cropping in the space of multiple class boundaries was considered. Results showed that the method ensured uniform distribution of samples and reduced the time of classification. Therefore, under the Weka platform, the improved KNN algorithm is effective.

Echo state network contains a randomly connected hidden layer and an adaptable output layer. It can overcome the problems associated with the complex computation and local optima. But there may be ill-posed problem when large reservoir state matrix is used to calculate the output weights by least square estimation. In this study, we use L1/2 regularization to calculate the output weights to get a sparse solution in order to solve the ill-posed problem and improve the generalized performance. In addition, an operation of iterated prediction is conducted to test the effectiveness of the proposed L1/2ESN for capturing the dynamics of the chaotic time series. Experimental results illustrate that the predictor has been designed properly. It outperforms other modified ESN models in both sparsity and accuracy.

The crowdsourcing and internet of things (IoT) have played a significant role in revolutionizing the information age. There is significant literature which has attempted to investigate the role of crowdsourcing and IoT in improving disaster response. However there is hardly any literature which attempted to reflect upon integration of crowdsourcing and IoT to explain better coordination among disaster relief workers to improve disaster response. In response to pressing need, we have attempted to develop a theoretical framework which can help disaster relief workers to improve their coordination using valuable information derived using comprehensive crowdsourcing framework. In this study we have used two-prong research strategies. First we have conducted extensive review of articles published in reputable journals, magazines and blogs by eminent practitioners and policy makers followed by case studies: stampede in Godavari River at Rajahmundry (2015), earthquake in Nepal (2015), flood in Uttarakhand (2013). Finally we have concluded our research findings with further research directions.

In this paper, we present an echo state network model based on sparse Gaussian process regression, which has been successfully applied to multivariate time series prediction. While combining the Gaussian process with Echo State Network, the computational complexity of the model is very high. We consider using a group of limited basis functions instead of the original covariance function, which reduces the computational complexity and maintains the prediction performance of the model. In the framework of Bayesian inference, the model can combine prior knowledge and observation data perfectly and provide prediction confidence. The model realizes adaptive estimation of the hyper-parameters by using maximum likelihood approach and avoids complex computation process. Two simulation results show the effectiveness and practicality of the proposed method.

Echo state network is a novel recurrent neural network, with a fixed reservoir structure and an adaptable linear readout layer, facilitating the application of RNNs. Often the network works beautifully. But sometimes it works poorly because of ill-posed problem. To solve it, we introduce a new approach towards ESNs, termed FSDESN, herein. It combines the merits of ESNs and fast subspace decomposition algorithm to provide a more precise alternative to conventional ESNs. The basic idea is to extract the subspace of the redundant large-scale reservoir state matrix by Krylov subspace decomposition algorithm, subsequently, calculate the readout weights using the subspace to replace original reservoir matrix. Hence, it can eliminate approximate collinear components and overcome the ill-posed problems so as to improve generalization performance. We exhibit the merits of our model in two multivariate benchmark datasets. Experimental results substantiate the effectiveness and characteristics of FSDESN.