Ling Cen

Linear Aperiodic Array Synthesis Using an Improved Genetic Algorithm

A novel algorithm on beam pattern synthesis for linear aperiodic arrays with arbitrary geometrical configuration is presented in this paper. Linear aperiodic arrays are attractive for their advantages on higher spatial resolution and lower sidelobe. However, the advantages are attained at the cost of solving a complex non-linear optimization problem. In this paper, we explain the Improved Genetic Algorithm (IGA) that simultaneously adjusts the weight coefficients and inter-sensor spacings of a linear aperiodic array in more details and extend the investigations to include the effects of mutual coupling and the sensitivity of the Peak Sidelobe Level (PSL) to steering angles. Numerical results show that the PSL of the synthesized beam pattern has been successfully lowered with the IGA when compared with other techniques published in the literature. In addition, the computational cost of our algorithm can be as low as 10% of that of a recently reported genetic algorithm based synthesis method. The excellent performance of IGA makes it a promising optimization algorithm where expensive cost functions are involved.

Linear Sparse Array Synthesis With Minimum Number of Sensors

The number of sensors employed in an array affects the array performance, computational load, and cost. Consequently, the minimization of the number of sensors is of great importance in practice. However, relatively fewer research works have been reported on the later. In this paper, a novel optimization method is proposed to address this issue. In the proposed method, the improved genetic algorithm that has been presented at a conference recently, is used to optimize the weight coefficients and sensor positions of the array. Sensors that contribute the least to the array performance are then removed systematically until the smallest acceptable number of sensors is obtained. Specifically, this paper reports the study on the relationship between the peak sidelobe level and the sensor weights, and uses the later to select the sensors to be removed. Through this approach, the desired beam pattern can be synthesized using the smallest number of sensors efficiently. Numerical results show that the proposed sensor removal method is able to achieve good sidelobe suppression with a smaller number of sensors compared to other existing algorithms. The computational load required by our proposed approach is about one order less than that required by other existing algorithms too.

An improved genetic algorithm for aperiodic array synthesis

In this paper, a novel algorithm on beam pattern synthesis for linear aperiodic arrays with arbitrary geometrical configuration is proposed. The algorithm is based on an improved genetic algorithm (IGA) that simultaneously adjusts the weight coefficients and inter-sensor spacings of a linear aperiodic array. A novel section-based crossover and a self-supervised mutation process are developed to improve the convergence performance. The results from simulation illustrate that with the IGA, the peak sidelobe level (PSL) of the synthesized beam pattern has been successfully lowered. In addition, the computational cost of the proposed algorithm can be as low as being about 10% of that of a recently reported genetic algorithm based synthesis method. The robustness of the proposed IGA has been illustrated clearly from the statistic of multiple independent runs too. The excellent performance of the IGA makes it a promising optimization algorithm where expensive cost functions are involved.

Linear sparse array synthesis via convex optimization

Due to the constraint on half-wavelength inter-element spacing of a uniformly spaced array, sparse arrays are usually designed to be non-uniformly spaced. Through proper design, unequally-spaced sparse arrays can have higher spatial resolution, lower sidelobe and less number of sensors required in comparison with uniformly spaced arrays. However, the synthesis of a sparse array is a non-convex process as the array beampattern is an exponential or trigonometric function of sensor positions. In this paper, we propose a synthesis scheme, where the design of sparse arrays is formulated as a convex optimization problem. The array weights are optimized to achieve minimum Peak Sidelobe Level (PSL) as well as maximizing the sparsity of the array by optimizing an objective function that includes two terms, one measures the PSL and the other measures the sparsity of array. Sparse array is then obtained by removing those sensors with weights approximately equal to zero. The proposed design scheme eliminates the need of optimization according to sensor positions, which consequently solves the problem in non-convex optimization that cannot guarantee to find the optimum solution with reasonable computation time. Numerical studies show that it can be successfully applied to sparse array synthesis with low computational complexity. Moreover, lower PSL and higher resolution of mainlobe can be achieved in comparison with a uniformly spaced array.

A Hybrid PNN-GMM classification scheme for speech emotion recognition

With the increasing demand for spoken language interfaces in human-computer interactions, automatic recognition of emotional states from human speeches has become of increasing importance. In this paper, we propose a novel hybrid scheme that combines the probabilistic neural network (PNN) and the Gaussian mixture model (GMM) for identifying emotions from speech signals. In order to handle mismatches more effectively, the universal background model (UBM) is incorporated into the GMM, and the resultant model is denoted as UBM-GMM. In the hybrid scheme, the strengths of the PNN and the UBM-GMM are combined through a novel conditional-probability based fusion algorithm. Experimental results show that the proposed scheme is able to achieve higher recognition accuracy than that obtained by using PNN or UBM-GMM alone.

Speech Emotion Recognition Using Canonical Correlation Analysis and Probabilistic Neural Network

In this paper, automatic identification of emotional states from human speech is addressed. While several papers have been published in the literature on speech emotion recognition, the features used are taken or modified from those used for speech recognition purposes. However, not all features used for speech recognition are of equal importance for emotion recognition. This paper addresses this issue and proposes a systematic method on feature selection for emotion recognition from speech signals. The idea is to work on a well-selected small feature set and use it to remove irrelevant information. Specifically, the proposed method uses the similar idea of the Canonical Correlation Analysis (CCA) to estimate the linear relationship between the various features and the emotional states. The outcome is a set of features that are of most relevance to the emotions. Experiments have been conducted using the LDC database and with the use of the Probabilistic Neural Network (PNN) as the classification method. The results obtained show that, comparable accuracies can be obtained for the emotional states tested with the use of only about 30% of the features considered. This implies that the computational load can be reduced greatly too.

Automatic Recognition of Emotional States From Human Speeches

As computer-based applications receive increasing attention in the real world and in our daily life, the Human-Computer Interaction (HCI) technology has also advanced rapidly over the recent decades. One essential enabler of natural interaction between human and computers is the computers’ ability to understand the emotional states expressed by the human subjects (so that personalized responses can be delivered accordingly). This is not surprising as it is well known that emotion plays an important role in human-human communications. Emotions are mental and physiological states associated with feelings, thoughts, and behaviors of human subjects. The emotional state expressed by a human subject reflects not only the mood but also the personality of the human subject. Automatic recognition of emotional states from cues expressed by human subjects, such as face expression or tone of voice, has found increasing applications in security, learning, medicine, entertainment, etc. For example, detecting abnormal emotions, such as stress or nervousness, helps to detect lie or identify suspicious human subjects. Emotion recognition in automatic tutoring systems, such as web-based e-learning, can adjust the tutoring content and delivery speed according to users’ responses. Automatically recognizing emotions from patients could also be helpful in clinical studies as well as in psychosis monitoring and diagnosis assistance. Emotion classification has been applied in the customer service sector

Deep Learning Method for Sleep Stage Classification

When humans fall asleep, they go through five sleep stages, i.e. wakefulness, stages of non-rapid eye movement consisting of N1, N2 and N3, and rapid eye movement (REM). Monitoring the proportion and distribution of sleep stages can help to diagnose sleep disorder and measure sleep quality. Traditional process of sleep scoring by well-trained experts is quite subjective and time-consuming. Automatic sleep staging analysis has demonstrated a lot of usefulness and attracted increasing attentions. With the massively growing size of accessible data and the rapid development of computational power, Deep Learning (DL) has achieved significant improvement in a lot of areas. In this work, an intelligent system for sleep stage classification is developed by using polysomnographic (PSG) data including electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG) based on a DL architecture. In our method, the Convolutional Neural Network (CNN) is employed as the feature detector, which is combined with a Hidden Markov Model (HMM) for its strengths of dealing with temporal data. Experiment results have shown a performance improvement compared to those methods with hand-crafted features or unsupervised feature learning by Deep Brief Learning (DBN).