Time-frequency Components Research Articles

Intra-pulse modulation discrimination using a self-supervised attention-driven CNN-BiLSTM-VAE combination

Identification of intra-pulse modulation (IPM) of radar signals is a crucial part of contemporary electronic support systems and electronic intelligence reconnaissance. Artificial intelligence (AI)-based methods can be very effective in recognising the IPM of radar signals. In this direction, an automatic method is proposed for recognising a few IPMs of radar signals based on continuous wavelet transform (CWT) and a hybrid model of self-attention (SA)-aided convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM). Firstly, time–frequency attributes of different radar signals are obtained using CWT, and thereafter CNN-SA-BiLSTM is utilised for feature extraction from the 2D scalograms formed by the time–frequency components. The CNN extracts features from the scalograms, SA enhances the discriminative power of the feature map, and BiLSTM detects radar signals based on these features. Additionally, the study addresses real-world data imbalance issues by incorporating a generative AI model, namely the Variational Autoencoder (VAE). The VAE-based approach effectively mitigates challenges arising from data imbalance situations. This method is tested at varying noise levels to give a proper representation of the actual electronic warfare environment. The simulation results demonstrate that the best overall recognition accuracy of the proposed method is 98.4%, even at low signal-to-noise ratios (SNR).

Rail crack identification of in-service turnout through Bayesian inference

In-service railway turnout, suffering from multi-factor coupling of operational and environmental loadings and special wheel–rail interaction, is prone to structural damage. The paper aims to develop a Bayesian damage identification method for crack-alike damage of the turnout rail under uncertainties. It consists of three core parts, including a damage index (DI) constructed by the transformation of time–frequency components of responses for generating damage-sensitive relationships, Bayesian models describing the crack-sensitive relationships hidden in the members of the damage index, and a mechanism synthesizing individual assessment results associated with the different reference models for providing one quantitative solution. The abnormal change in the stable relationships derived from the index can reflect damage occurrence and severity. The Bayesian approach is adopted to model the relationships under uncertainties of in-service railway turnout. The models trained by using monitoring data of the turnout rail before being damaged serve as a reference for healthy state, and deviations of actual observations from model predictions may indicate the existence of damage. The synthesizing process helps to offer a more rational assessment result through the weighted summation of individual quantitative assessment results. Rail monitoring data of a railway turnout are acquired to examine the damage detection performance of the proposed method. By exempting loadings measurement and physical model derivation, this data-driven methodology is potentially capable of supporting the damage identification and quantitative assessment of other structures in railway engineering.

Deep neural network based time–frequency decomposition for structural seismic responses training with synthetic samples

AbstractTime–frequency decomposition is a powerful tool in assessing the dynamic behaviors of structures. Traditional time–frequency decomposition methods struggle with adaptability, and are limited in handling the structural responses with strong nonlinearity and closely spaced modes. In this study, a cutting‐edge approach based on deep neural network (DNN) is proposed to achieve a precise and adaptive time–frequency decomposition of nonlinear structural responses under seismic excitations. Remarkably, despite being trained on synthetic samples, the proposed method demonstrates outstanding performance in decomposing time–frequency components across various seismic response scenarios. Compared to variational mode decomposition (VMD) and synchroextracting transform (SET), the proposed method exhibited superior precision in time–frequency decomposition and excellent efficiency in parameter optimization. Moreover, the applicability of the proposed method to real‐world complex structures has been verified, which also shows promising generalization capabilities. Future research will aim to enhance the network performance by incorporating additional learning samples with diverse nonlinear characteristics.

The impact of geopolitical risks, financial stress, economic policy uncertainty on African stock markets returns and volatilities: wavelet coherence analysis

PurposeThis study seeks to examine the relationship between macroeconomic shock indicators, namely geopolitical risk (GPR), global economic policy uncertainty (GEPU) and financial stress (FS), and returns as well as volatilities on seven carefully selected stock markets in Africa. Specifically, the study intends to unravel the co-movement and interdependence between the respective macroeconomic shock indicators and each of the stock markets under consideration across time and frequency.Design/methodology/approachThis study employed wavelet coherence approach to examine the strength and stability of the relationships across different time scales and frequency components, thereby providing valuable insights into specific periods and frequency ranges where the relationships are particularly pronounced.FindingsThe study found that GEPU, Financial Stress (FS) and GPR failed to induce significant influence on African stock market returns in the short term (0–4 months band), but tend to intensify in the long-term band (after 6th month). On the contrary, stock market volatilities exhibited strong coherence and interdependence with GEPU, FSI and GPR in the short-term band.Originality/valueThis study happens to be the first of its kind to comprehensively consider how the aforementioned macro-economic shock indicators impact stock markets returns and volatilities over time and frequency. Further, none of the earlier studies has attempted to examine the relationship between macro-economic shocks, stock returns and volatilities in different crisis periods. This study is the first of its kind in to employ data spanning from May 2007 to April 2023, thereby covering notable crisis periods such as global financial crisis (GFC) and the COVID-19 pandemic episodes.

Heightened neural reward responsiveness functions as a plasticity factor moderating the association between childhood emotional abuse and young adult depressive symptoms: Evidence of differential susceptibility.

As a neural indicator of reward responsiveness (RR), reward positivity (RewP) has been demonstrated to moderate the association between stress exposure and depressive symptoms. However, extant research has primarily (a) focused on life stress rather than early maltreatment, (b) ignored the time-frequency components, and (c) has been based on a traditional perspective of diathesis stress. The present study aimed to comprehensively examine whether and how neurophysiological (RewP and its time-frequency decomposition components) and self-reported measures of RR interact with childhood emotional abuse on young adult depressive symptoms. The sample of 192 Chinese university students aged 18-25 (Mage = 21.08 ± 1.91 years; 59.4% girls) completed self-reported questionnaires of emotional abuse, depressive symptoms and RR. The RewP and its time-frequency components delta and theta were elicited via a monetary reward task. The results demonstrated that RewP significantly moderated the association between emotional abuse and young adult depressive symptoms in a differential susceptibility but not diathesis-stress manner. However, gain-related delta, loss-related theta, or self-reported RR did not drive such moderation effects. These findings were robust and survived a series of rigorous sensitivity analyses. The current findings provide preliminary evidence that heightened RewP may function as a plasticity factor moderating the association between early maltreatment exposure and depression, and highlight the effect specific to emotional abuse. However, caution should be paid to the generalizability of these findings in high-risk clinical samples, in light of the current high-functioning sample features and low rates of high symptom and abuse levels. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Machine Fault Diagnosis through Vibration Analysis: Continuous Wavelet Transform with Complex Morlet Wavelet and Time–Frequency RGB Image Recognition via Convolutional Neural Network

In pursuit of advancing fault diagnosis in electromechanical systems, this research focusses on vibration analysis through innovative techniques. The study unfolds in a structured manner, beginning with an introduction that situates the research question in a broader context, emphasising the critical role of fault diagnosis. Subsequently, the methods section offers a concise summary of the primary techniques employed, highlighting the utilisation of short-time Fourier transform (STFT) and continuous wavelet transform (CWT) for extracting time–frequency components from the signal. The results section succinctly summarises the main findings of the article, showcasing the results of features extraction by CWT and subsequently utilising a convolutional neural network (CNN) for fault diagnosis. The proposed method, named CWTx6-CNN, was compared with the STFTx6-CNN method of the previous stage of the investigation. Visual insights into the time–frequency characteristics of the inertial measurement unit (IMU) data are presented for various operational classes, offering a clear representation of fault-related features. Finally, the conclusion section underscores the advantages of the suggested method, particularly the concentration of single-frequency components for enhanced fault representation. The research demonstrates commendable classification performance, highlighting the efficiency of the suggested approach in real-time scenarios of fault analysis in less than 50 ms. Calculation by CWT with a complex Morlet wavelet of six time–frequency images and combining them into a single colour image took less than 35 ms. In this study, interpretability techniques have been employed to address the imperative need for transparency in intricate neural network models, particularly in the context of the case presented. Notably, techniques such as Grad-CAM (gradient-weighted class activation mapping), occlusion, and LIME (locally interpretable model-agnostic explanation) have proven instrumental in elucidating the inner workings of the model. Through a comparative analysis of the proposed CWTx6-CNN method and the reference STFTx6-CNN method, the application of interpretability techniques, including Grad-CAM, occlusion, and LIME, has played a pivotal role in revealing the distinctive spectral representations of these methodologies.

Time-Frequency Image-based Speech Emotion Recognition using Artificial Neural Network

Automatic Speech Emotion Recognition (ASER) is a state-of-the-art application in artificial intelligence. Speech recognition intelligence is employed in various applications such as digital assistance, security, and other human-machine interactive products. In the present work, three open-source acoustic datasets, namely SAVEE, RAVDESS, and EmoDB, have been utilized (Haq et al., 2008, Livingstone et al., 2005, Burkhardt et al., 2005). From these datasets, six emotions namely anger, disgust, fear, happy, neutral, and sad, are selected for automatic speech emotion recognition. Various types of algorithms are already reported for extracting emotional content from acoustic signals. This work proposes a time-frequency (t-f) image-based multiclass speech emotion classification model for the six emotions mentioned above. The proposed model extracts 472 grayscale image features from the t-f images of speech signals. The t-f image is a visual representation of the time component and frequency component at that time in the two-dimensional space, and differing colors show its amplitude. An artificial neural network-based multiclass machine learning approach is used to classify selected emotions. The experimental results show that the above-mentioned emotions' average classification accuracy (CA) of 88.6%, 85.5%, and 93.56% is achieved using SAVEE, RAVDESS, and EmoDB datasets, respectively. Also, an average CA of 83.44% has been achieved for the combination of all three datasets. The maximum reported average classification accuracy (CA) using spectrogram for SAVEE, RAVDESS, and EmoDB dataset is 87.8%, 79.5 %, and 83.4%, respectively (Wani et al., 2020, Mustaqeem and Kwon, 2019, Badshah et al., 2017). The proposed t-f image-based classification model shows improvement in average CA by 0.91%, 7.54%, and 12.18 % for SAVEE, RAVDESS, and EmoDB datasets, respectively. This study can be helpful in human-computer interface applications to detect emotions precisely from acoustic signals.

Damage Identification of Turnout Rail through a Covariance-Based Condition Index and Quantitative Pattern Analysis

Subjected to complex loadings from the wheel–rail interaction, turnout rail is prone to crack damage. This paper aims to develop a condition evaluation method for crack-alike damage detection of in-service turnout rail. A covariance-based structural condition index (CI) is firstly constructed by fusing the time-frequency components of responses, generating a series of patterns governed by the interrelationships between column members in the CI matrix. The damage-sensitive interrelationships latent in CI are then modeled using Bayesian regression and historical data, and baseline patterns are built with predictions of the models and new inputs. The deviations between the baseline patterns and the actual patterns of the newly observed CI members are quantitatively assessed. To synthetically consider the individual assessment results, a technique is developed to combine the individual assessment results into one synthetic result by designing a group of suitable weights taking into consideration both probabilistic confidence and reference model error. If the deviations are within a tolerable range, no damage is flagged; otherwise, damage existence and severity are reported. A case study is conducted, in which monitoring data from the database of a railway turnout are applied to build the CI matrix and examine the damage identification performance of this method. Good agreement between actual conditions and assessment results is found in different testing scenarios in the case study, demonstrating the effectiveness of the proposed method.

Attention driven CWT-deep learning approach for discrimination of Radar PRI modulation

With the proliferation of radio frequency (RF) systems and radar applications, Electronic Warfare (EW) is receiving increasing importance. The analysis of the radar signals is a critical EW task that decides the nature of counter employments. In an Electronic Support (ES) system, the challenge is to detect hostile radiation sources efficiently and trigger a counter response. Detection of types of Pulse Repetition Interval (PRI) modulation of radar signal significantly facilitates the manifestation of RF emitters during recognition which is difficult in a dense EW environment. Recent developments in artificial intelligence (AI) methods suggest that this emerging technology can be effective for such purposes. In this direction, an automatic approach for recognizing several kinds of complex PRI modulation based on Continuous Wavelet Transform (CWT) and a combination of the vanilla Convolutional Neural Network (CNN), a multi-head self-attention (MHSA) mechanism and the popular Long Short-Term Memory (LSTM) is proposed. The CWT is used to decompose the PRI modulation sequence and obtain different time–frequency components. Further, aided by the proposed CNN-MHSA-LSTM combination, the features extracted from the CWT 2D-scalograms are used to execute PRI modulation discrimination. In this method, the vanilla CNN is employed for the extraction of deep features to figure out the class details while capturing the spatial attributes. Thereafter, to improve the discriminative power of the entire framework a MHSA mechanism is used. The temporal attributes are acquired by the LSTM which works in concert with the CNN for executing the detection of the PRI classes based on the extracted features. Also to assess the effectiveness of the proposed method, three models based on ResNet, popular CNN and SqueezeNet are implemented for benchmark comparison in terms of overall performance and complexity. The simulation results show that the proposed method enhances performance and achieves robustness in the noise-filled and imperfect channel knowledge environment. The best recognition accuracy is 98.3% with 50% spurious pulses in the environment which fluctuates with imperfect channel knowledge cases.

Deep learning model to identify potential occurrence of reduced LVEF in patients with implantable loop recorders

Abstract Background Early detection of left ventricular systolic dysfunction (LVSD) using minimally invasive electrocardiographic assessment in an ambulatory setting may be important considering the increasing economic burden of HF. Recently artificial intelligence (AI) algorithms have been reported to classify reduced ejection fraction using snapshot 12-lead ECG measurements, however the ability of AI to identify LVSD using only a single lead ambulatory ECG is unknown. Purpose We aimed to develop a convolution neural network (CNN) based deep learning algorithm using single lead ECG to predict ejection fraction less than 40% in ambulatory patients implanted with implantable loop recorders (ILRs). Methods We linked ILR patients with LVEF measurements from a de-identified database of aggregated electronic health record (EHR) data during the period of 2007-2021 to a manufacturer’s device database with ambulatory single lead ECGS. The routine ECG transmissions from the ILR devices were paired with LVEF measured within a week. The data was pre-processed to obtain the time-frequency components of ECG using wavelet transform. The wavelet coefficients were then used as input to train and validate a 2D-CNN model. An independent test dataset consisting of data from patient not included in training/validation dataset was used to assess the performance of CNN model to classify LVEF< = 40% using the metrics area under the curve (AUC), sensitivity and specificity. Results A total of 35,741 unique LVEF-ECG pairs collected from 2,249 ILR patients were used to train the model; an independent validation dataset consisted of 6,721 unique LVEF ECG pairs from 750 patients and independent test dataset consisted of 6,611 unique LVEF ECG pairs from 750 patients. In our CNN model, cross entropy was used as loss function and adaptive moment estimation (ADAM) was used as the optimizer. The initial learning rate was selected as 0.0001, epochs as 20 and mini batch size as 64. The threshold to delineate reduced ejection fraction (LVEF< = 40%) was set as 0.1 (Figure 1). The model yielded accuracy, sensitivity, specificity, AUC of 75%,70%, 76% and 0.8 respectively in the independent test set (Figure 2). Conclusion A deep learning algorithm applied to ambulatory single lead ECGs acquired by ILR can detect LVEF< = 40%. This continuous ambulatory AI-ECG monitoring may allow to identify longitudinal changes in LVEF post ILR implant via remote monitoring.Figure 1:Model output vs Actual EF rangeFigure 2:ROC

Hilbert–Huang Transformation-based subject-specific time–frequency-space pattern optimization for motor imagery electroencephalogram classification

The advancement of brain–computer interfaces (BCIs) has narrowed the gap between humans and computers, allowing intentional interaction by monitoring and translating brain signals in real time. Among BCI approaches, motor imagery electroencephalogram (MI-EEG) systems are popular due to their non-invasiveness, portability, and user-friendly operation without external stimuli. However, MI-EEG classification faces challenges from subject-specific variations in time, frequency, and spatial domains. To overcome this, the paper proposes a novel Hilbert–Huang transformation (HHT)-based method for subject-specific time–frequency-space pattern optimization in MI-EEG classification. The method utilizes a joint time–frequency pattern optimization module and a spatial pattern optimization module for EEG measurements. This efficient process identifies subject-specific dominant time–frequency components and extracts optimal spatial features. The optimized features are fed into a support vector machine (SVM) classifier, resulting in superior performance compared to standard baselines on three open-source datasets. The proposed method achieves 4.1% and 6.3% higher accuracy in 2-class and 4-class classification, respectively. Additionally, it demonstrates remarkable computational efficiency, requiring 70% less training time to achieve the optimal feature space. These improvements in classification accuracy and computational efficiency underscore the practical value of the proposed method for MI-BCI systems.

Time-frequency component of the GreenX library: minimax grids for efficient RPA and GW calculations

Time-frequency component of the GreenX library: minimax grids for efficient RPA and GW calculations

Intelligent Fault Diagnosis of Robotic Strain Wave Gear Reducer Using Area-Metric-Based Sampling

The main challenge with rotating machine fault diagnosis is the condition monitoring of machines undergoing nonstationary operations. One possible way of efficiently handling this situation is to use the deep learning (DL) method. However, most DL methods have difficulties when the issue of imbalanced datasets occurs. This paper proposes a novel framework to mitigate this issue by developing an area-metric-based sampling method. In the proposed process, the new sampling scheme can identify which locations of the datasets can potentially have a high degree of surprise. The basic idea of the proposed method is whenever significant deviations from the area metrics are observed to populate more sample points. In addition, to improve the training accuracy of the DL method, the obtained sampled datasets are transformed into a continuous wavelet transform (CWT)-based scalogram representing the time–frequency component. The dilated convolutional neural network (CNN) is also introduced as a classification process with the altered images. The efficacy of the proposed method is demonstrated with fault diagnosis problems for welding robots. The obtained results are also compared with existing methods.

Time Frequency Analysis-Based Averaging and Fusion of Features for Wearable Non-Invasive Blood Glucose Estimation

This paper proposes a Clark error grid based stacking method for fusing various time frequency averaged features together for reducing the measurement errors for performing the wearable non-invasive blood glucose estimation. Like the conventional time frequency representation based denoising methods, the most critical issue is on the selection of the time frequency components. In this paper, three time frequency representation based averaging methods are investigated. They are the discrete cosine transform (DCT) based averaging method, the singular spectrum analysis (SSA) based averaging method and the empirical mode decomposition (EMD) based averaging method. It is found that retaining 40% of the DCT coefficients as well as three SSA components and three intrinsic mode functions (IMFs) could yield the best blood glucose estimation performances. Two datasets are employed for demonstrating the effectiveness and the robustness of our proposed method. The most common blood glucose estimation method and our proposed method can yield the Pearson’s correlation coefficient (R) at 0.4308 and 0.9172, respectively, the mean absolute error (MAE) at 0.8783 and 0.8193, respectively, the root mean squares error (RMSE) at 1.0680 and 0.9621, respectively, and the mean absolute relative difference (MARD) at 0.1504 and 0.1133, respectively, as well as the percentage of the data falling in the zone A of the Clark error grid at 71.1111% and 90.5882%, respectively, for the data in the first dataset. The results obtained for the data in the second dataset are similar. Obviously, our proposed method significantly outperforms the most common blood glucose estimation methods.

Automatic bolt tightness detection using acoustic emission and deep learning

Bolts are typically deployed to unify disparate components in civil engineering structures, making the surveillance of their tightness critical to ensuring the stability of such structures. The innovation of this paper is the development of an automatic monitoring method for bolt tightness using acoustic emission (AE) and deep learning (DL). Initially, AE sensors are first used to monitor structural bolts, and AE signals are preprocessed by a continuous wavelet transform (CWT) to obtain different time–frequency components, which are then fed into the convolutional neural network (CNN). To reduce the amount of training parameters and enhance the capacity for feature extraction, transfer learning (TL) is introduced to train the model. In the experimental scenario, the bolts exhibit seven categories of tightness. The findings show that the proposed method can accurately classify the bolts’ tightness, and can be used to monitor the early looseness of the bolts online.

Identifying Intraoperative Spinal Cord Injury Location from Somatosensory Evoked Potentials' Time-Frequency Components.

Excessive distraction in corrective spine surgery can lead to iatrogenic distraction spinal cord injury. Diagnosis of the location of the spinal cord injury helps in early removal of the injury source. The time-frequency components of the somatosensory evoked potential have been reported to provide information on the location of spinal cord injury, but most studies have focused on contusion injuries of the cervical spine. In this study, we established 19 rat models of distraction spinal cord injury at different levels and collected the somatosensory evoked potentials of the hindlimb and extracted their time-frequency components. Subsequently, we used k-medoid clustering and naive Bayes to classify spinal cord injury at the C5 and C6 level, as well as spinal cord injury at the cervical, thoracic, and lumbar spine, respectively. The results showed that there was a significant delay in the latency of the time-frequency components distributed between 15 and 30 ms and 50 and 150 Hz in all spinal cord injury groups. The overall classification accuracy was 88.28% and 84.87%. The results demonstrate that the k-medoid clustering and naive Bayes methods are capable of extracting the time-frequency component information depending on the spinal cord injury location and suggest that the somatosensory evoked potential has the potential to diagnose the location of a spinal cord injury.

TS-LSSVM: Triple sparse least squares support vector machine for residual oxygen concentration detection of encapsulated pharmaceutical vials

Accurate measurement of residual oxygen concentration in encapsulated pharmaceutical vials is adequate to ensure the quality of inner sterile preparations. However, the critical characteristic signal is feeble and covered by enormous environmental interference in the actual production. Inspired by structural sparse learning, we propose a novel prediction model in this paper, triple sparse least squares support vector machine (TS-LSSVM), in which the production priors are deeply excavated, and the feature, sample, and structure sparsity are realized simultaneously by redefining the objective function. In addition, the selection of support vectors can be adjusted adaptively according to the time-varying environmental noise, so as to ensure the reliability of long-term operation. First, the time–frequency components containing the prior knowledge are extracted based on the synchro squeezing wavelet transform (SSWT). Then, a triple sparse learning strategy is designed, which can accurately eliminate redundant wavelet coefficients and adaptively select training samples. Finally, a strict and fast optimization solution is proposed under the alternating direction method of multipliers (ADMM) framework. Experimental results on public and practical datasets prove the superiority of TS-LSSVM.

A failure detection method based on multivariate variational mode decomposition and skewness for modular DC circuit breakers

Modular DC circuit breakers (MDCCBs) are widely used to protect multi-terminal HVDC (MT-HVDC) girds. However, the metal-oxide varistor (MOV) failure detection of MDCCBs still poses a challenge for the reliability of MT-HVDC grids. To this end, this paper analyzes the faulty features of the MDCCB and proposes a failure detection method that uses a multivariate variational mode decomposition (MVMD) and a differential skewness. The proposed method is as follows. First, MVMD decomposes the differential current, which is obtained by the current of the MDCCB energy absorption branch, into its time–frequency components, and then the time–frequency component that has the highest frequency is selected as the feature component. Second, a differential skewness is constructed by the feature component, and the differential skewness is employed to detect the MOV failure of the MDCCB. Finally, the performance of the proposed method is evaluated by a four-terminal symmetric bipolar HVDC system with MDCCBs. The simulation experiment results show that the proposed method can correctly detect the MOV failures of the MDCCB in different faulty conditions, even in the weak failure condition with noise influence.

Multiple classification of EEG signals and epileptic seizure diagnosis with combined deep learning

Epilepsy stands out as one of the common neurological diseases. The neural activity of the brain is observed using electroencephalography (EEG), which allows the diagnosis of epilepsy disease. The aim of this study is to create a combined deep learning model that automatically detects epileptic seizure activity, detection of the epileptic region and classifies EEG signals by using images representing the time-frequency components of the time series EEG signal and numerical values of the raw EEG signals. In the study, 3 different public datasets, CHB-MIT, Bern-Barcelona and Bonn EEG records were used. This study presents a combined model using the time sequence of EEG signals and time-frequency-image transformations of time-dependent EEG signals. CWT and STFT methods were used to convert signals to images. Two models were created separately with the images created by CWT and STFT methods. In the Bonn dataset average accuracy rates of 99.07 %, 99.28 %, respectively, in binary classifications and 97.60 % and 98.56 %, respectively, in multiple classifications were obtained with scalogram and spectrogram images. In the Bern-Barcelona and CHB-MIT datasets, 95.46 % and 96.23 % accuracy rates were obtained, respectively. The data combinations brought together in 3 different combinations with the Bonn dataset were underwent to 8-fold cross validation and average accuracy rates of 99.21 % (± 0.56), 99.50 % (± 0.45), and 98.84 % (± 1.58) were obtained. The model we created can detect whether there is epileptic seizure activity in EEG data, detection of the epileptic region and classify EEG signals with a high success rate.

Time-Frequency Component-Aware Convolutional Neural Network for Wireless Interference Classification

In wireless communication systems, wireless interference classification (WIC) is considered as one of the most effective technologies to address the challenges brought by electromagnetic interference in military and civilian scenarios. Recently, deep learning (DL) based methods have dominated progress in the field of WIC. However, the most existing methods do not consider the redundancy of input samples, nor have the ability to adaptively allocate computational resources conditioned on the inputs. To this end, we propose time-frequency component-aware convolutional neural network (TFCCNN), and it allows the convolution calculation to be performed only at the locations where time-frequency components or important parts exist in the time-frequency image of interference signals, leading to reduce the superfluous computation. Furthermore, to further reduce the computational complexity, we introduce a novel adaptive forward propagation (AFP) algorithm, and the network can determine the depth of forward propagation according to the difficulty of the sample during inference. Experimental results demonstrate that the proposed method reduces the computational complexity by about 75% when the recognition accuracy is slightly improved compared to the traditional CNNs.