Local Discriminant Bases Research Articles

Objective classification of auditory brainstem responses to consonant-vowel syllables using local discriminant bases

ObjectiveThe main goal of this study is classifying the auditory brainstem responses to consonant-vowels /da/, /ba/, and /ga/ into three classes automatically. MethodAuditory brainstem responses (ABRs) to consonant-vowel syllables /da/, /ba/, and /ga/ were recorded from twenty-seven normal subjects. Time domain features (i.e., energy, entropy, cosine distance and correlation coefficient) and frequency domain features (i.e., magnitude and phase of the frequency responses) were extracted. Also, local discriminant bases (LDB) method was employed to extract time-frequency features from wavelet packet coefficients. Then, random subset feature selection (RSFS) algorithm reduced the feature spaces. Afterward, discriminant analysis (DA), naive Bayes (NB), multiclass support vector machine (MSVM) and K-nearest neighbors (KNN) methods classified the selected features. ResultsThe time–frequency domain features showed better results than other features (91.36%). The maximum accuracy was achieved for MSVM classifier when we used a combination of all the features (97.5%). ConclusionThis study shows the efficiency of frequency and time–frequency domains features. The results indicate that time–frequency features obtained by local discriminant bases were more successful in objective classifications of the responses. Besides, the selected features from the phase of the frequency responses were reliable in classifying the responses to consonant-vowels /da/, /ba/ and /ga/. SignificanceThe importance of this study lies in the fact that it helps the objective classification of auditory brainstem responses underlying the different encoding of three consonant-vowels (/ba/, /da/ and /ga/) in frequency and time–frequency domains.

Band Selection Using Forward Feature Selection Algorithm for Citrus Huanglongbing Disease Detection

, 2015Purpose: This study investigated different band selection methods to classify spectrally similar data - obtained from aerial images of healthy citrus canopies and citrus greening disease (Huanglongbing or HLB) infected canopies - using small differences without unmixing endmember components and therefore without the need for an endmember library. However, large number of hyperspectral bands has high redundancy which had to be reduced through band selection. The objective, therefore, was to first select the best set of bands and then detect citrus Huanglongbing infected canopies using these bands in aerial hyperspectral images. Methods: The forward feature selection algorithm (FFSA) was chosen for band selection. The selected bands were used for identifying HLB infected pixels using various classifiers such as K nearest neighbor (KNN), support vector machine (SVM), naive Bayesian classifier (NBC), and generalized local discriminant bases (LDB). All bands were also utilized to compare results. Results: It was determined that a few well-chosen bands yielded much better results than when all bands were chosen, and brought the classification results on par with standard hyperspectral classification techniques such as spectral angle mapper (SAM) and mixture tuned matched filtering (MTMF). Median detection accuracies ranged from 66-80%, which showed great potential toward rapid detection of the disease. Conclusions: Among the methods investigated, a support vector machine classifier combined with the forward feature selection algorithm yielded the best results.Keywords: Bayesian classification, Hyperspectral, K nearest neighbor, Multi-modal Bayesian classification, Support vector machine

Machine Fault Classification Based on Local Discriminant Bases and Locality Preserving Projections

Machine fault classification is an important task for intelligent identification of the health patterns for a mechanical system being monitored. Effective feature extraction of vibration data is very critical to reliable classification of machine faults with different types and severities. In this paper, a new method is proposed to acquire the sensitive features through a combination of local discriminant bases (LDB) and locality preserving projections (LPP). In the method, the LDB is employed to select the optimal wavelet packet (WP) nodes that exhibit high discrimination from a redundant WP library of wavelet packet transform (WPT). Considering that the obtained discriminatory features on these selected nodes characterize the class pattern in different sensitivity, the LPP is then applied to address mining inherent class pattern feature embedded in the raw features. The proposed feature extraction method combines the merits of LDB and LPP and extracts the inherent pattern structure embedded in the discriminatory feature values of samples in different classes. Therefore, the proposed feature not only considers the discriminatory features themselves but also considers the dynamic sensitive class pattern structure. The effectiveness of the proposed feature is verified by case studies on vibration data-based classification of bearing fault types and severities.

Aero-Engine Fault Diagnosis Using Improved Local Discriminant Bases and Support Vector Machine

This paper presents an effective approach for aero-engine fault diagnosis with focus on rub-impact, through combination of improved local discriminant bases (LDB) with support vector machine (SVM). The improved LDB algorithm, using both the normalized energy difference and the relative entropy as quantification measures, is applied to choose the optimal set of orthogonal subspaces for wavelet packet transform- (WPT-) based signal decomposition. Then two optimal sets of orthogonal subspaces have been obtained and the energy features extracted from those subspaces appearing in both sets will be selected as input to a SVM classifier to diagnose aero-engine faults. Experiment studies conducted on an aero-engine rub-impact test system have verified the effectiveness of the proposed approach for classifying working conditions of aero-engines.

Extracting optimal tempo-spatial features using local discriminant bases and common spatial patterns for brain computer interfacing

Brain computer interfaces (BCI) provide a new approach to human computer communication, where the control is realised via performing mental tasks such as motor imagery (MI). In this study, we investigate a novel method to automatically segment electroencephalographic (EEG) data within a trial and extract features accordingly in order to improve the performance of MI data classification techniques. A new local discriminant bases (LDB) algorithm using common spatial patterns (CSP) projection as transform function is proposed for automatic trial segmentation. CSP is also used for feature extraction following trial segmentation. This new technique also allows to obtain a more accurate picture of the most relevant temporal–spatial points in the EEG during the MI. The results are compared with other standard temporal segmentation techniques such as sliding window and LDB based on the local cosine transform (LCT).

Noninvasive detection of mechanical prosthetic heart valve disorder

Auscultation is a widely used efficient technique by cardiologists for detecting the heart conditions. Since the mechanical prosthetic heart valves are widely used today, it is important to develop a simple and efficient method to detect abnormal mechanical valves. In this paper, the mechanical prosthetic heart valve sounds are analyzed by using different power spectral density (PSD) estimation techniques. To improve the classification accuracy of heart sounds, we propose two different feature extraction schemes, i.e., a modified local discriminant bases (LDB) scheme and a Hilbert–Huang Transform (HHT)-based scheme. A database of 150 heart sounds is used in this study and an average classification accuracy of 97.3% is achieved for both the two feature extraction schemes, when a generic linear discriminant analysis (LDA) classifier is used in the classification stage.

Determination of steel quality based on discriminating textural feature selection

Image data can be acquired from a product surface in real time by image sensor systems in chemical plants. For quality determination based on these image datasets, effective texture classification methodology is essential to handle highly dimensional images and to extract quality-related information from these product surface images. Wavelet texture analysis is useful for reducing the dimension and extracting textural information from images. Although wavelet texture analysis extracts only textural characteristics from images, the extracted features still contain unnecessary information for classification. The texture analysis method can be improved by retaining only class-dependent features and removing common features. In previous works, best basis and local discriminant basis are the most popular techniques for selecting an important basis from the wavelet packet basis. However, feature selection based on wavelet texture analysis has been studied for texture classification. Because previous methods are designed for wavelet coefficients with features for analysis, their performance is poor with wavelet texture analysis. We propose a novel texture classification methodology for quality determination based on feature selection using wavelet texture analysis. The proposed methodology applies the sequential forward floating selection (SFFS) algorithm as a feature selection strategy to select discriminating wavelet signatures using wavelet texture analysis. The proposed methodology is validated through quality determination for industrial steel surfaces. The results show that the proposed method has fewer classification errors with fewer number of features than previous methods.

Information Theory Filters for Wavelet Packet Coefficient Selection with Application to Corrosion Type Identification from Acoustic Emission Signals

The damage caused by corrosion in chemical process installations can lead to unexpected plant shutdowns and the leakage of potentially toxic chemicals into the environment. When subjected to corrosion, structural changes in the material occur, leading to energy releases as acoustic waves. This acoustic activity can in turn be used for corrosion monitoring, and even for predicting the type of corrosion. Here we apply wavelet packet decomposition to extract features from acoustic emission signals. We then use the extracted wavelet packet coefficients for distinguishing between the most important types of corrosion processes in the chemical process industry: uniform corrosion, pitting and stress corrosion cracking. The local discriminant basis selection algorithm can be considered as a standard for the selection of the most discriminative wavelet coefficients. However, it does not take the statistical dependencies between wavelet coefficients into account. We show that, when these dependencies are ignored, a lower accuracy is obtained in predicting the corrosion type. We compare several mutual information filters to take these dependencies into account in order to arrive at a more accurate prediction.

Genetic algorithm for informative basis function selection from the wavelet packet decomposition with application to corrosion identification using acoustic emission

Chemical process installations are exposed to aggressive chemicals and conditions leading to corrosion. The damage from corrosion can lead to an unexpected plant shutdown and to the exposure of people and the environment to chemicals. Due to changes within and on the surface of materials subjected to corrosion, energy is released in the form of acoustic waves. This acoustic activity can be captured and used for corrosion monitoring in chemical process installations. Wavelet packet coefficients extracted from the acoustic activity have been considered to determine whether corrosion occurs, and to identify the type of corrosion process, at least for the most important corrosion processes in the chemical process industry. Feature subset selection is then applied to these wavelet coefficients to achieve a much higher accuracy in the identification of different corrosion processes than when no feature subset selection is applied to the acoustic waves. However, due to the statistical dependencies that potentially exist between the wavelet coefficients, the latter should not be selected independently from each other. Local discriminant basis selection algorithms do not take the statistical dependencies between wavelet coefficients into account. In this paper, we have used several mutual information-based approaches that take these dependencies into account and compared them to the wavelet-specific local discriminant basis selection algorithm. Furthermore, a hybrid filter–wrapper genetic algorithm, which uses a relevance–redundancy approach as a local search procedure, was designed. The highest classification accuracies are obtained with the hybrid filter–wrapper genetic algorithm, for all classifiers used in this paper. Furthermore, the proposed algorithm easily outperformed one of the most commonly used classifiers in chemometrics: partial least squares discriminant analysis (PLS-DA). A naïve Bayes classifier that uses the features selected by the hybrid filter–wrapper genetic algorithm was able to identify the absence of corrosion, uniform corrosion, pitting and stress corrosion cracking, with an accuracy of up to 87.20%.

Detection of contaminated hazelnuts and ground red chili pepper flakes by multispectral imaging

Mycotoxins are the toxic metabolites of certain filamentous fungi and have been demonstrated to cause various health problems in humans, including immunosuppression and cancer. Among them, the aflatoxins have received greater attention because they are potent carcinogens and are responsible for many human deaths per annum, mostly in non-industrialized countries. Various regulatory agencies have enforced limits on the concentrations of these toxins in foods and feeds involved in international commerce. Hyperspectral and multispectral imaging are becoming increasingly important for rapid and nondestructive testing for the presence of such contaminants. However, the high number of spectral bands needed may render such image acquisition systems too complex, expensive and slow. Moreover, they tend to generate overwhelming amount of data, making effective processing of this information in real time difficult. In this study, a two-dimensional local discriminant bases algorithm was developed to detect the location of the discriminative features in the multispectral data space. The algorithm identifies the optimal passband width and center frequencies of optical filters to be used for a multispectral imaging system. This was applied to a multispectral imaging system used to detect aflatoxin-contaminated hazelnut kernels and red chili peppers. Classification accuracies of 92.3% and 80% were achieved for aflatoxin-contaminated and uncontaminated hazelnuts and red chili peppers, respectively. The aflatoxin concentrations were decreased from 608 to 0.84ppb for tested hazelnuts and from 38.26 to 22.85ppb for red chili peppers by removal of the nuts/peppers that were classified as aflatoxin-contaminated. The algorithm was also used to classify fungal contaminated and uncontaminated hazelnut kernels, and an accuracy of 95.6% was achieved for this broader classification.

WAVELET PACKET DECOMPOSITION FOR THE IDENTIFICATION OF CORROSION TYPE FROM ACOUSTIC EMISSION SIGNALS

Corrosion causes a degradation of the structural integrity of petrochemical plants, nuclear power plants, ships, bridges and other constructions containing steel with the consequence that people and the environment may be exposed to dangerous situations. The detection of corrosion and the prediction of the type of corrosion are studied in this article by means of the acoustic emission technique. We use a wavelet packet decomposition to compute features from the acoustic emission signals. The basis functions with the highest discriminative power are selected according to the highest pair-wise Kullback–Leibler divergence between distributions of wavelet coefficients. It is proven that the pair-wise Kullback–Leibler divergence used in the local discriminant basis algorithm requires class conditional independence of the wavelet coefficients. Several classification algorithms using the most discriminative wavelet coefficients are compared for the prediction of three types of corrosion and the absence of corrosion.

Use of supervised discretization with PCA in wavelet packet transformation-based surface electromyogram classification

This paper describes a preprocessing stage for nonlinear classifier used in wavelet packet transformation (WPT)-based multichannel surface electromyogram (EMG) classification. The preprocessing stage named sdPCA, which consists of supervised discretization coupled with principal component analysis (PCA), was developed for improving surface EMG classifier generalization ability and training speed on overlap segmented signals. The sdPCA outperforms the fast correlation-based filter (FCBF), PCA, supervised discretization, and their combinations in terms of the highest generalization ability, fast training speed, the small feature size, and an ability to reduce the risks of developing oscillation and being trapped in nonlinear classifier training. The experiments were conducted on a data set consisting of 4-channel surface EMG signals measured from 6 hand and wrist gestures of 12 subjects. The experimental results indicate that the classification system using sdPCA has the highest generalization ability along with the second fastest training speed. The classification accuracy in 12 subjects of the system using sdPCA is 93.30 ± 2.42% taking 400 epochs for training by overlap segmented signals within 100 s. This result is very attractive for further development because we can achieve high-classification accuracy for large data sets by means of the proposed sdPCA without the application of additional algorithms such as local discriminant bases (LDB), majority voting (MV), or WPT sub-bands clustering.

Breast Tumor Characterization Based on Ultrawideband Microwave Backscatter

Characterization of architectural tissue features such as the shape, margin, and size of a suspicious lesion is commonly performed in conjunction with medical imaging to provide clues about the nature of an abnormality. In this paper, we numerically investigate the feasibility of using multichannel microwave backscatter in the 1-11 GHz band to classify the salient features of a dielectric target. We consider targets with three shape characteristics: smooth, microlobulated, and spiculated; and four size categories ranging from 0.5 to 2 cm in diameter. The numerical target constructs are based on Gaussian random spheres allowing for moderate shape irregularities. We perform shape and size classification for a range of signal-to-noise ratios (SNRs) to demonstrate the potential for tumor characterization based on ultrawideband (UWB) microwave backscatter. We approach classification with two basis selection methods from the literature: local discriminant bases and principal component analysis. Using these methods, we construct linear classifiers where a subset of the bases expansion vectors are the input features and we evaluate the average rate of correct classification as a performance measure. We demonstrate that for 10 dB SNR, the target size is very reliably classified with over 97% accuracy averaged over 360 targets; target shape is classified with over 70% accuracy. The relationship between the SNR of the test data and classifier performance is also explored. The results of this study are very encouraging and suggest that both shape and size characteristics of a dielectric target can be classified directly from its UWB backscatter. Hence, characterization can easily be performed in conjunction with UWB radar-based breast cancer detection without requiring any special hardware or additional data collection.

Identification of Damaged Wheat Kernels and Cracked-Shell Hazelnuts with Impact Acoustics Time-Frequency Patterns

A new adaptive time-frequency (t-f) analysis and classification procedure is applied to impact acoustic signals for detecting hazelnuts with cracked shells and three types of damaged wheat kernels. Kernels were dropped onto a steel plate, and the resulting impact acoustic signals were recorded with a PC-based data acquisition system. These signals were segmented with a flexible local discriminant bases (F-LDB) procedure in the time-frequency plane to extract discriminative patterns between damaged and undamaged food kernels. The F-LDB procedure requires no prior knowledge of the relevant time or frequency indices of the impact acoustics signals for classification. The method automatically finds all crucial time-frequency indices from the training data by combining local cosine packet analysis and a frequency axis clustering approach, which supports individual time and frequency band adaptation. Discriminant features are extracted from the adaptively segmented acoustic signal, sorted according to a Fisher class separability criterion, post-processed by principal component analysis, and fed to a linear discriminant classifier. Experimental results establish the superior performance of the proposed approach when compared to prior techniques reported in the literature or used in the field. The new approach separated damaged wheat kernels (IDK, pupal, and scab) from undamaged wheat kernels with 96%, 82%, and 94% accuracy, respectively. It also separated cracked-shell hazelnuts from those with undamaged shells with 97.1% accuracy. The adaptation capability of the algorithm to the time-frequency patterns of signals makes it a universal method for food kernel inspection that can resist the impact acoustic variability between different kernel and damage types.

Audio Signal Feature Extraction and Classification Using Local Discriminant Bases

Audio feature extraction plays an important role in analyzing and characterizing audio content. Auditory scene analysis, content-based retrieval, indexing, and fingerprinting of audio are few of the applications that require efficient feature extraction. The key to extract strong features that characterize the complex nature of audio signals is to identify their discriminatory subspaces. In this paper, we propose an audio feature extraction and a multigroup classification scheme that focuses on identifying discriminatory time-frequency subspaces using the local discriminant bases (LDB) technique. Two dissimilarity measures were used in the process of selecting the LDB nodes and extracting features from them. The extracted features were then fed to a linear discriminant analysis-based classifier for a three-level hierarchical classification of audio signals into ten classes. In the first level, the audio signals were grouped into artificial and natural sounds. Each of the first level groups were subdivided to form the second level groups viz. instrumental, automobile, human, and nonhuman sounds. The third level was formed by subdividing the four groups of the second level into the final ten groups (drums, flute, piano, aircraft, helicopter, male, female, animals, birds and insects). A database of 213 audio signals were used in this study and an average classification accuracy of 83% for the first level (113 artificial and 100 natural sounds), 92% for the second level (73 instrumental and 40 automobile sounds; 40 human and 60 nonhuman sounds), and 89% for the third level (27 drums, 15 flute, and 31 piano sounds; 23 aircraft and 17 helicopter sounds; 20 male and 20 female speech; 20 animals, 20 birds and 20 insects sounds) were achieved. In addition to the above, a separate classification was also performed combining the LDB features with the mel-frequency cepstral coefficients. The average classification accuracies achieved using the combined features were 91% for the first level, 99% for the second level, and 95% for the third level

Time-Width Versus Frequency Band Mapping of Energy Distributions

Most of the signal processing operations involve some kind of a transformation or approximation of the signal. The transform coefficients or the approximation parameters reveal many hidden characteristics of a signal that can be appropriately processed to extract useful information. In recent years, adaptive time-frequency (TF) transformations have significantly contributed to this area. The TF transformation can be classified into two main categories based on 1) signal decomposition approaches and 2) bilinear TF distributions (also known as Cohen's class). TF distributions are nonparametric in nature and mainly used for visualization purposes. On the other hand, decomposition approaches are parametric in nature and highly suitable for objective feature extraction. This paper focuses on a particular TF decomposition approach (adaptive TF transformation) based on matching pursuit-type algorithm. Using this decomposition approach, a novel time-width versus frequency band (TWFB) energy mapping is proposed that possesses both parameterization benefits and meaningful visual patterns with favorable properties for pattern recognition. This organized mapping of the TF decomposition parameters allows the application of pruning algorithms such as local discriminant bases (LDB) to identify application specific TF subspaces. The identification of these subspaces enables efficient processing of information and reduces the computational effort considerably. The visual patterns of the TWFB mappings exhibit high potential of becoming a powerful pattern analysis tool. The paper covers in detail the formulation of the TWFB mapping and some of its properties. Experiments performed with speech and synthetic signals produced desirable results demonstrating the benefits of TWFB for pattern recognition related applications

A comparison of model and non-model based time-frequency transforms for sperm whale click classification

We tried to find discriminating features for sperm whale clicks in order to distinguish between clicks from different whales, or to enable unique identification. We examined two different methods to obtain suitable characteristics. First, a model based on the Gabor function was used to describe the dominant frequencies in a click, and then the model parameters were used as classification features. The Gabor function model was selected because it has been used to model dolphin sonar pulses with great precision. Additionally, it has the interesting property that it has an optimal time–frequency resolution. As such, it can indicate optimal usage of the sonar by sperm whales. Second, the clicks were expressed in a wavelet packet table, from which subsequently a local discriminant basis was created. A wavelet packet basis has the advantage that it offers a highly redundant number of coefficients, which allow signals to be represented in many different ways. From the redundant signal description a representation can be selected that emphasizes the differences between classes. This local discriminant basis is more flexible than the Gabor function, which can make it more suitable for classification, but it is also more complex. Class vectors were created with both models and classification was based on the distance of a click to these vectors. We show that the Gabor function could not model the sperm whale clicks very well, due to the variability of the changing click characteristics. Best performance was reached when three subsequent clicks were averaged to smoothen the variability. Around 70% of the clicks classified correctly in both the training and validation sets. The wavelet packet table adapted better to the changing characteristics, and gave better classification. Here, also using a 3-click moving average, around 95% of the training sets classified correctly and 78% of the validation sets. These numbers lowered by only a few per cent when single clicks, instead of a moving average, were classified. This indicates that, while the features may show too much variability to enable unique identification of individual whales on a click by click basis, the wavelet approach may be capable of distinguishing between a small group of whales.

Neural network-based sperm whale click classification

Recordings of a group of foraging sperm whales usually result in a mixture of clicks from different animals. To analyse the click sequences of individual whales these clicks need to be separated, and for this an automatic classifier would be preferred. Here we study the use of a radial basis function network to perform the separation. The neural network's ability to discriminate between different whales was tested with six data sets of individually diving males. The data consisted of five shorter click trains and one complete dive which was especially important to evaluate the capacity of the network to generalize. The network was trained with characteristics extracted from the six click series with the help of a wavelet packet-based local discriminant basis. The selected features were separated in a training set containing 50 clicks of each data set and a validation set with the remaining clicks. After the network was trained it could correctly classify around 90% of the short click series, while for the entire dive this percentage was around 78%.

Classification of single trial motor imagery EEG recordings with subject adapted non-dyadic arbitrary time–frequency tilings

We describe a new technique for the classification of motor imagery electroencephalogram (EEG) recordings in a brain computer interface (BCI) task. The technique is based on an adaptive time–frequency analysis of EEG signals computed using local discriminant bases (LDB) derived from local cosine packets (LCP). In an offline step, the EEG data obtained from the C3/C4 electrode locations of the standard 10/20 system is adaptively segmented in time, over a non-dyadic grid by maximizing the probabilistic distances between expansion coefficients corresponding to left and right hand movement imagery. This is followed by a frequency domain clustering procedure in each adapted time segment to maximize the discrimination power of the resulting time–frequency features. Then, the most discriminant features from the resulting arbitrarily segmented time-frequency plane are sorted. A principal component analysis (PCA) step is applied to reduce the dimensionality of the feature space. This reduced feature set is finally fed to a linear discriminant for classification. The online step simply computes the reduced dimensionality features determined by the offline step and feeds them to the linear discriminant. We provide experimental data to show that the method can adapt to physio-anatomical differences, subject-specific and hemisphere-specific motor imagery patterns. The algorithm was applied to all nine subjects of the BCI Competition 2002. The classification performance of the proposed algorithm varied between 70% and 92.6% across subjects using just two electrodes. The average classification accuracy was 80.6%. For comparison, we also implemented an adaptive autoregressive model based classification procedure that achieved an average error rate of 76.3% on the same subjects, and higher error rates than the proposed approach on each individual subject.

Fault classification on vibration data with wavelet based feature selection scheme

Fault classification based upon vibration measurements is an essential building block of a conditional based health usage monitoring system. Multiple sensors are incorporated to assure the redundancy and to achieve the desired reliability and accuracy. The shortcoming of using multiple sensors is the need to deal with a high dimensional feature set, a computationally expensive task in classification. It is vital to reduce the feature dimension via an effective feature extraction and feature selection algorithm. A simple wavelet based feature selection scheme is proposed herein, uniquely built by local discriminant bases and genetic optimization. This scheme overcomes the disadvantages faced by the existing feature selection methods by producing a generic feature set, reducing the dimensionality of features, and requiring no prior information of the problem domain. The proposed feature selection scheme is based upon the strategy of “divide and conquer” that significantly reduce the computation time without compromising the classification performance. The simulation results show the proposed feature selection scheme provides at least 65% reduction of the total number of features at no cost of the classification accuracy.