Independent component analysis (ICA) is a well known technique of blind source separation (BSS) and is used in various applications, e.g. speech, biomedical, communication, robotics, leakage detection, vibration analysis and machinery fault diagnosis. The ICA technique estimates the original source signals from the recorded multidimensional mixed signals through various sensors from any physical process. In case one or more sensor becomes faulty the separation of mixed signals becomes very difficult. In machinery, ICA is used to diagnose faults in its rotating parts that is a major concern of public safety. In case of faulty sensors, fault diagnosis becomes difficult. Moreover, in certain situations of wireless sensor networks some of the sensors fail to collect accurate information. Therefore, to collect accurate information in case of sensor failure, fault diagnosis technique is required. In this paper, a sensor fault diagnosis technique called the state observing technique (SOT) is developed to first diagnose faults in the system and then identify the faulty sensors. Also the extended sensor technique (EST) is developed to improve the separation performance of the ICA algorithm in case of faulty sensors based on information provided by the SOT technique. Effectiveness of the proposed SOT–EST technique is evaluated through extensive simulations utilizing the FastICA algorithm of ICA. To the best of our knowledge, we are the first to discuss the sensor fault diagnosis in the ICA applications.

ICA Based Sensors Fault Diagnosis: An Audio Separation Application

Application and comparison of feature-based classification models for multistable impact motions of percussive drilling

Deep Learning Based Low Complexity Symbol Detection and Modulation Classification Detector

Large penetration of renewable energy sources into the power grid has increased the complexity of power system operation and greatly reduced the ability of the system to withstand large disturbances. The frequent occurrence of voltage instability problems in the power grid has brought new challenges to the assessment of transient stability analysis of the power system. To achieve fast and accurate assessment of transient voltages, a transfer learning-based transient voltage stability assessment with small samples is proposed, introducing a domain transfer learning approach, embedding a cooperative attention mechanism in the residual network during the feature extraction stage to capture long-range correlations between features, and using adversarial approaches to reduce the differences between samples from different data sets, using the source domain to guide the target domain for network training to improve model’s evaluation capability when the number of samples is insufficient, enhance the generalisation performance of the network, and effectively improve the performance of real-time power system transient voltage stability evaluation in the absence of sufficient historical data. Testing on an improved New England 39-node system validates the superiority of this method in transient voltage stability assessment and provides a new approach to practical field transient voltage stability assessment.

Transient voltage stability assessment based on transfer learning for small samples

In maritime communications systems, there are marine VHF communications systems that meet GMDSS standards in addition to AIS and VDES systems which use very high-frequency signals for information communications such as security rescue. But because its communication does not contain the identity information, the channel is easy to be occupied maliciously, thus interferes in normal maritime communication. This paper studies and analyzes the individual identification technology of the VHF signal based on the rf fingerprint technology of signal. Using the improved adversarial generation network ACWGAN(Auxiliary Classifier Wasserstein Generative Adversarial Networks) to train and identify, we obtain a better classification result. The recognition rate can reach 94% when the SNR is 5 dB for 10 different classes of VHF signal.

An Intelligent Maritime Communication Signal Recognition Algorithm based on ACWGAN

Modulation classification is the technological basis of adaptive interference mitigation in communication system. This paper proposes a modulation classification method for very high frequency (VHF) signals, which is based on deep convolutional neural network (CNN) and cyclic spectrum graphs. First, the cyclic spectrum of VHF signals is analyzed. Then, a deep learning method based on CNN is proposed, down-sampling and clipping technologies are used for preprocessing cyclic spectrum images, parameters of the proposed neural network are optimized, and finally the modulation classification is realized. The experimental results show that, the proposed method has high modulation classification accuracy and less computation burden in low SNR.

Modulation Classification of VHF Communication System based on CNN and Cyclic Spectrum Graphs

The anti-jamming of communication radio has attracted intensive attentions from communication community. The blind source separation (BSS) is much effective when the jamming has entered. The multiple kernel independent component analysis (MKICA) is proposed instead of a single one, which can not only combine multiple kernels corresponding to different notions of similarity or information from multiple feature subsets, but also fuse distinctions of multiple kernels. The core of algorithm is the kernel canonical correlation analysis (KCCA), where the efficient use of multiple kernel trick makes it more suitable for different sources and more robust for different distributions. Hence, there will be a remarkable improvement of separation accuracy. Numerical experiments based on the artificially synthesized data and intermediate frequency (IF) signal containing clear voice demonstrate state-of-the-art performances of proposed algorithm. It is ideal for separation of useful communication signals when confronting with various jamming in the anti-jamming field.

Multiple Kernel Independent Component Analysis for Anti-jamming of Communication Radio

In this paper, we introduce a channel discrepancies adaptive automatic modulation recognition (AMR) method, which employs the domain adversarial training (DAT) to tackle the issue of wireless channel mismatch between training and testing conditions. The channel mismatch is a critical problem for deep learning (DL) based AMR systems. In realistic scenarios, the channel environment mismatch commonly happens and a large mismatch may seriously degrade the recognition accuracy of signals. The introduced channel discrepancies adaptive AMR method consists of a l-dimensional convolutional neural network (1-D CNN) based recognition model and a domain discriminator model. The DAT encourages the 1-D CNN to extract channel invariant features and increase the robustness of the AMR system to new channel environment. We evaluate the proposed method and competition approaches on the popular RadioML2016. 04c and RadioML2016.10a dataset. Experimental results shows that the introduced channel discrepancies adaptive AMR system produce notable better recognition performance than that of the methods without domain adaptation for the channel discrepancies of training and testing datasets.

Channel Discrepancies Adaptive Modulation Recognition Using Domain Adversarial Training

In the anti-jamming field of radio communication, the problem of online and multiclass jamming recognition is fundamental to implement reasonable anti-jamming measures. The incremental small sphere and large margin (IncSSLM) is proposed, this model can learn the compact boundary for own communication signals and known jamming, which relieves the open-set problem of radio data. Meanwhile it can also update the model of classifier in real time, which avoids the large memory requirement for vast jamming data and saving much time for training. The core of proposed method is the small sphere and large margin (SSLM) approach, which makes the spherical area as compact as possible, like support vector data description (SVDD), and also makes the margin between them as far as possible, like support vector machine (SVM). In other words, it can minimize intra-class divergence and maximize inter-class space. Therefore, there is a significant enhancement of recognition performance when compared with open classifiers such as SVM, and considerable superiority of training efficiency when compared with the canonical SSLM algorithm. Numerical experiments based on synthetic data, practical complex feature data of high-resolution range profile (HRRP), and jamming data of radio communication demonstrate that IncSSLM is efficient and promising for multiple and online recognition of vase and open-set radio jamming.

Guo Yu

Papers

Modulation Classification of VHF Communication System based on CNN and Cyclic Spectrum Graphs

Multiple Kernel Independent Component Analysis for Anti-jamming of Communication Radio

Channel Discrepancies Adaptive Modulation Recognition Using Domain Adversarial Training

Incremental small sphere and large margin for online recognition of communication jamming

Single and multiple frame coding of LSF parameters using deep neural network and pyramid vector quantizer