Waveform Imaging Research Articles

Alzheimer's disease recognition based on waveform and spectral speech signal processing.

Alzheimer's disease (AD) is a neurodegenerative disorder with an irreversible progression. Currently, it is diagnosed using invasive and costly methods, such as cerebrospinal fluid analysis, neuroimaging, and neuropsychological assessments. Recent studies indicate that certain changes in language ability can predict early cognitive decline, highlighting the potential of speech analysis in AD recognition. Based on this premise, this study proposes an AD recognition multi-channel network framework, which is referred to as the ADNet. It integrates both time-domain and frequency-domain features of speech signals, using waveform images and log-Mel spectrograms derived from raw speech as data sources. The framework employs inverted residual blocks to enhance the learning of low-level time-domain features and uses gated multi-information units to effectively combine local and global frequency-domain features. The study tests it on a dataset from the Shanghai cognitive screening (SCS) digital neuropsychological assessment. The results show that the method we proposed outperforms existing speech-based methods, achieving an accuracy of 88.57%, a precision of 88.67%, and a recall of 88.64%. This study demonstrates that the proposed framework can effectively distinguish between the AD and normal controls, and it may be useful for developing early recognition tools for AD.

Internal Quality Inspection and Characterization of Casting Billets Based on Ultrasonic Microscopy

Abstract For evaluating the internal quality of casting billets, the low-power experimental and metallographic methods have the disadvantages of the complex process, the small detection range and the destructive evaluation. In addition, the internal defects of the casting billet are analyzed by the one-dimensional waveform and two-dimensional image for the traditional ultrasonic testing, which can not describe directly the three-dimensional spatial information. Hence, a new method is proposed based on ultrasonic microscopy to detect and characterize the internal defection of the casting billet. In this paper, 42CrMo casting billet is regarded as objective, five sets of samples are fabricated by adjusting temperature gradient field and the number of presses. The ultrasonic sequence images of the present samples are achieved by using ultrasonic microscopy, and the spatial distributions, number, and sizes are obtained for the defects of the casting billet by reconstructing the three-dimensional appearance of microdefects to characterize the internal mass of the casting billet. Finally, the effective and accuracy of the proposed method are validated by using metallographic method.

Artificial intelligence enabled interpretation of ECG images to predict hematopoietic cell transplantation toxicity

AbstractArtificial intelligence (AI)–enabled interpretation of electrocardiogram (ECG) images (AI-ECGs) can identify patterns predictive of future adverse cardiac events. We hypothesized that such an approach would provide prognostic information for the risk of cardiac complications and mortality in patients undergoing hematopoietic cell transplantation (HCT). We retrospectively subjected ECGs obtained before HCT to an externally trained, deep-learning model designed to predict the risk of atrial fibrillation (AF). Included were 1377 patients (849 autologous [auto] HCT and 528 allogeneic [allo] HCT recipients). The median follow-up was 2.9 years. The 3-year cumulative incidence of AF was 9% (95% confidence interval [CI], 7-12) in patients who underwent auto HCT and 13% (10%-16%) in patients who underwent allo HCT. In the entire cohort, pre-HCT AI-ECG estimate of AF risk correlated highly with the development of clinical AF (hazard ratio [HR], 7.37; 95% CI, 3.53-15.4; P < .001), inferior survival (HR, 2.4; 95% CI, 1.3-4.5; P = .004), and greater risk of nonrelapse mortality (NRM; HR, 95% CI, 3.36; 1.39-8.13; P = .007), without increased risk of relapse. Association with mortality was only noted in allo HCT recipients, where the risk of NRM was greater. The use of cyclophosphamide after transplantation resulted in greater 90-day incidence of AF (13% vs 5%; P = .01) compared to calcineurin inhibitor–based graft-versus-host disease prophylaxis, corresponding to temporal changes in AI-ECG AF prediction after HCT. In summary, AI-ECG can inform risk of posttransplant cardiac outcomes and survival in HCT patients and represents a novel strategy for personalized risk assessment.

Deep Learning-based Low Frequency Extrapolation: Its implication in 2D Full Waveform Imaging for Marine Seismic Data in the Sadewa Field, Indonesia

Deep Learning-based Low Frequency Extrapolation: Its implication in 2D Full Waveform Imaging for Marine Seismic Data in the Sadewa Field, Indonesia

Cuffless blood pressure estimation from photoplethysmography using deep convolutional neural network and transfer learning

Blood pressure (BP) monitoring is an essential indicator for diseases of the cardiovascular system. Early detection of abnormalities in BP helps to significantly reduce the risk of diseases such as chronic heart failure, stroke and hypertension. In this study, we propose a new framework for systolic, diastolic and mean arterial blood pressure (SBP, DBP and MBP) estimation using PPG signals and its derivatives. The proposed framework includes the extraction of deep features by pre-trained CNN models based on transfer learning using images of signal waveforms. The most distinctive features are obtained by using the RFE feature selection algorithm on the extracted deep features. The selected deep features are exposed to a range of well-known machine and deep learning regression methods. The proposed BP estimation models have been evaluated with statistical metrics, visual analytical tools and international gold standards such as AAMI and BHS. Experimental results show that the first derivative of PPG, the VPG input image, the deep features obtained with DenseNet121 and the bi-directional gated recurrent unit (Bi-GRU) algorithm provide the best BP estimation performance. The results reveal that the SBP, DBP and MBP estimation models are grade-A according to the BHS protocol and meet the AAMI standard. The presented framework has been compared with well-established studies using the MIMIC II dataset, and the comparative analysis confirms that the proposed method outperforms existing techniques. The deep learning model for BP estimation offers a highly accurate, fast and practical system, especially in the monitoring of patients at risk of hypertension.

Deep-learning-based renal artery stenosis diagnosis via multimodal fusion.

Diagnosing Renal artery stenosis (RAS) presents challenges. This research aimed to develop a deep learning model for the computer-aided diagnosis of RAS, utilizing multimodal fusion technology based on ultrasound scanning images, spectral waveforms, and clinical information. A total of 1485 patients received renal artery ultrasonography from Peking Union Medical College Hospital were included and their color doppler sonography (CDS) images were classified according to anatomical site and left-right orientation. The RAS diagnosis was modeled as a process involving feature extraction and multimodal fusion. Three deep learning (DL) models (ResNeSt, ResNet, and XCiT) were trained on a multimodal dataset consisted of CDS images, spectrum waveform images, and individual basic information. Predicted performance of different models were compared with senior physician and evaluated on a test dataset (N=117 patients) with renal artery angiography results. Sample sizes of training and validation datasets were 3292 and 169 respectively. On test data (N=676 samples), predicted accuracies of three DL models were more than 80% and the ResNeSt achieved the accuracy 83.49%±0.45%, precision 81.89%±3.00%, and recall 76.97%±3.7%. There was no significant difference between the accuracy of ResNeSt and ResNet (82.84%±1.52%), and the ResNeSt was higher than the XCiT (80.71%±2.23%, p<0.05). Compared to the gold standard, renal artery angiography, the accuracy of ResNest model was 78.25%±1.62%, which was inferior to the senior physician (90.09%). Besides, compared to the multimodal fusion model, the performance of single-modal model on spectrum waveform images was relatively lower. The DL multimodal fusion model shows promising results in assisting RAS diagnosis.

Waveform Imaging Based on Linear Forward Representations for Scalar Wave Seismic Data

The current reverse-time migration, which is based on wave equations for imaging wavefields, employs an imaging formula derived from Claerbout’s imaging principle. This imaging formula is only valid for plane waves with small incident angles on the perfectly flat reflecting surface. However, the complexity of seismic wave propagation may lead to situations that do not meet this requirement. Therefore, this paper divides the subsurface into local scattering and reflecting bodies. It proposes linear forward representations for scattering and reflection data based on perturbations in the physical parameters and wave impedance, respectively. To further describe the effect on the reflecting body boundary, the local reflection coefficient is defined and the linear forward representation for the reflection data based on it is obtained. After that, the proposed linear forward representations are used as the forward equations for the linear inverse of the seismic data, and the seismic data waveform imaging method is developed based on linear inversion theory. At the same time, the specific waveform imaging calculation formulas for scalar wave scattering data and scalar wave reflection data are provided and validated via numerical experiments. Compared with the current reverse-time migration, waveform migration not only has the correct phase and higher resolution in theory but also does not increase the computational complexity. To some extent, it improves the deficiencies of the current structural imaging and provides a basis for subsurface lithological imaging.

Electrocardiographic deep learning for predicting post-procedural mortality: a model development and validation study

Preoperative risk assessments used in clinical practice are insufficient in their ability to identify risk for postoperative mortality. Deep-learning analysis of electrocardiography can identify hidden risk markers that can help to prognosticate postoperative mortality. We aimed to develop a prognostic model that accurately predicts postoperative mortality in patients undergoing medical procedures and who had received preoperative electrocardiographic diagnostic testing. In a derivation cohort of preoperative patients with available electrocardiograms (ECGs) from Cedars-Sinai Medical Center (Los Angeles, CA, USA) between Jan 1, 2015 and Dec 31, 2019, a deep-learning algorithm was developed to leverage waveform signals to discriminate postoperative mortality. We randomly split patients (8:1:1) into subsets for training, internal validation, and final algorithm test analyses. Model performance was assessed using area under the receiver operating characteristic curve (AUC) values in the hold-out test dataset and in two external hospital cohorts and compared with the established Revised Cardiac Risk Index (RCRI) score. The primary outcome was post-procedural mortality across three health-care systems. 45 969 patients had a complete ECG waveform image available for at least one 12-lead ECG performed within the 30 days before the procedure date (59 975 inpatient procedures and 112 794 ECGs): 36 839 patients in the training dataset, 4549 in the internal validation dataset, and 4581 in the internal test dataset. In the held-out internal test cohort, the algorithm discriminates mortality with an AUC value of 0·83 (95% CI 0·79-0·87), surpassing the discrimination of the RCRI score with an AUC of 0·67 (0·61-0·72). The algorithm similarly discriminated risk for mortality in two independent US health-care systems, with AUCs of 0·79 (0·75-0·83) and 0·75 (0·74-0·76), respectively. Patients determined to be high risk by the deep-learning model had an unadjusted odds ratio (OR) of 8·83 (5·57-13·20) for postoperative mortality compared with an unadjusted OR of 2·08 (0·77-3·50) for postoperative mortality for RCRI scores of more than 2. The deep-learning algorithm performed similarly for patients undergoing cardiac surgery (AUC 0·85 [0·77-0·92]), non-cardiac surgery (AUC 0·83 [0·79-0·88]), and catheterisation or endoscopy suite procedures (AUC 0·76 [0·72-0·81]). A deep-learning algorithm interpreting preoperative ECGs can improve discrimination of postoperative mortality. The deep-learning algorithm worked equally well for risk stratification of cardiac surgeries, non-cardiac surgeries, and catheterisation laboratory procedures, and was validated in three independent health-care systems. This algorithm can provide additional information to clinicians making the decision to perform medical procedures and stratify the risk of future complications. National Heart, Lung, and Blood Institute.

Microseismic Monitoring Signal Waveform Recognition and Classification: Review of Contemporary Techniques

Microseismic event identification is of great significance for enhancing our understanding of underground phenomena and ensuring geological safety. This paper employs a literature review approach to summarize the research progress on microseismic signal identification methods and techniques over the past decade. The advantages and limitations of commonly used identification methods are systematically analyzed and summarized. Extensive discussions have been conducted on cutting-edge machine learning models, such as convolutional neural networks (CNNs), and their applications in waveform image processing. These models exhibit the ability to automatically extract relevant features and achieve precise event classification, surpassing traditional methods. Building upon existing research, a comprehensive analysis of the strengths, weaknesses, opportunities, and threats (SWOT) of deep learning in microseismic event analysis is presented. While emphasizing the potential of deep learning techniques in microseismic event waveform image recognition and classification, we also acknowledge the future challenges associated with data availability, resource requirements, and specialized knowledge. As machine learning continues to advance, the integration of deep learning with microseismic analysis holds promise for advancing the monitoring and early warning of geological engineering disasters.

Cardiac interoception in Anorexia Nervosa: A resting-state heartbeat-evoked potential study.

A deficit in interoception - the ability to perceive, interpret and integrate afferent signals about the physiological state of the body - has been shown in Anorexia Nervosa (AN), and linked to altered hunger sensations, body dysmorphia, and abnormal emotional awareness. The present high-density electroencephalography (hdEEG) study aims to assess cardiac interoception in AN and to investigate its neural correlates, using an objective neurophysiological measure. Heartbeat-evoked potentials (HEPs) were computed from 5min of resting-state EEG and electrocardiogram (ECG) data and compared between individuals with AN (N=22) and healthy controls (HC) (N=19) with waveform, topographic, and source imaging analyses. Differences in the cortical representation of heartbeats were present between AN and HC at a time window of 332-348ms after the ECG R-peak. Source imaging analyses revealed a right-sided hypoactivation in AN of brain regions linked to interoceptive processing, such as the anterior cingulate and orbitofrontal areas. To the best of our knowledge, this is the first study using hdEEG to localise the underlying sources of HEPs in AN. Results point to altered interoceptive processing during resting-state in AN. As our participants had a short duration of illness, this might not be the consequence of prolonged starvation. Interventions targeted at interoception could provide an additional tool to facilitate recovery.

Characterization of 3D-Radar images of pavement devoid damage based on FDTD

Accurate judgement of devoid damage information by 3D-Radar is an effective way of repairing damage in nondestructive pavements. In order to systematically analyse the characteristics of devoid damage under nondestructive pavements in 3D-Radar response. In this study, the 3D-Radar response to devoid damage of different sizes, locations and moisture contents was quantified by FDTD orthorectified simulations. Data acquisition of the pre-buried devoid damage on site was carried out using 3D-Radar, compared with the orthorectified simulation results and numerical analysis. The detection effect was also verified by relying on the project. The results show that the radar wave characteristics of the devoid damage are obvious. Different colour and waveform image characteristics in B-Scan in the presence and absence of water at the location; the size of the devoid also has an impact on the image characteristics. It depends on the footprints and size of the devoid. It creates “upward-convex”, “down-concave” and straight features; the presence of the devoid characteristics in the 3D-Radar mapping will enhance the confidence of the devoid identification through field tests and engineering verification.

Comparison of size distributions of recovered particles generated from Ag wire in air by pulsed wire discharge and electric explosion of wire methods

Electric explosion of a wire (EEW) with discharge and pulsed wire discharge (PWD) are essential methods for producing nanosized and micronized metal particles. The study of Ag particle size distributions is important for Ag particle recovery from photovoltaic panel sheets using these methods performed in previous studies. In this study, we examined the size distributions of Ag particles generated from an Ag wire in air by EEW and PWD to clarify the difference between, and effects of these methods on particle generation. The particle size distribution generated in PWD was concentrated in a narrower range. The simulation results suggest that volume expansion occurred for a liquidized Ag wire in EEW and for a gasified Ag wire in PWD. Visualized images of processes and current waveforms in EEW and PWD indicated that after the wire explosion the restrike with plasma occurred in the expanding Ag particles in air in PWD. These results suggest that the narrower particle size distribution obtained in PWD may be attributable to monodisperse powderization by plasma evaporation of the gasified wire. These results show that PWD with plasma of an Ag wire in air effectively creates particles that are more monodispersed than those formed by EEW.

Recognition of shear and tension signals based on acoustic emission parameters and waveform using machine learning methods

Acoustic emission (AE) technology is widely used to monitor the damage evolution of rock. Identifying AE signals is crucial to reveal the rock cracking mechanism. The current tension and shear signal discrimination methods primarily rely on AE parameters. However, the differences between AE parameters and waveform images used as inputs require further exploration. Acoustic emission data from Brazilian splitting and direct shear tests of red sandstone were collected. A decision tree was applied to classify the signals based on the AE parameters: Rise time (RT), Counts, Amplitude, Peak frequency (PF), and absolute energy. Three typical CNN-based networks (InceptionV3, ResNet50, and VGG19) were used to classify the AE signals using images of waveforms, reconfigured waveforms, and spectrograms generated by wavelet packets. The results showed that using AE parameters as input to distinguish tensile and shear signals yields an accuracy of up to 88%, increasing recognition accuracy when using PF as input. In addition, increasing the number of input parameters to the decision tree can improve average recognition accuracy, with single, double, and three parameters being 65%, 78%, and 82%, respectively. However, when the number of parameters exceeds three, the accuracy improvement is minimal, with four and five parameters being 84% and 86%, respectively. The CNN-based networks outperform the decision tree in terms of recognition accuracy. It is recommended that the VGG19 network and waveform are adopted as input to identify the tension and shear AE signals. In contrast, using the spectrogram as input enhances the stability of the model, particularly for the validation set, although the improvement is limited.

Study on the penetration capability of GPR for the steel-fibre reinforced concrete (SFRC) segment based on numerical simulations and model test

The SFRC segment is gaining attention in shield tunnels due to their better mechanical properties. However, the existence of steel fibres poses a greater difficulty when attempting to utilize ground penetrating radar (GPR) for nondestructive quality inspection compared to conventional reinforced concrete segments. This paper presented a refined numerical model of SFRC segments with gprMax software, to investigate the penetration characteristics of GPR. The numerical model incorporated four different steel fibre contents, namely 0%, 0.5%, 2.5%, and 5%. The model also included an internal void and a surface metal plate to compare and analyze the waveform characteristics of GPR with the normal sections of the segment. The orthogonal test of numerical simulation investigates the imaging characteristics, power attenuation patterns, and imaging features of three different GPR frequency gradients: 200 MHz, 500 MHz, and 1000 MHz. This analysis helps in selecting the appropriate detection frequency based on the observed energy attenuation patterns and imaging characteristics. Then, the model test was conducted at the selected frequency to study the penetration performance of GPR. The research results indicate that the numerical model established in this paper demonstrates good effectiveness in studying the waveform characteristics of steel fibre-reinforced concrete under various conditions. The numerical model established in this paper realized the study of the penetration characteristics and imaging features of different steel fibre contents, GPR frequencies, and detection targets. Analysis of waveform characteristics, power attenuation, and imaging results indicates that GPR at 500 MHz exhibits better detection performance for SFRC segments. The results of the numerical simulation were verified through model tests of real SFRC segments, demonstrating that 500 MHz has good penetration performance for SFRC segments. This paper also proposed a calculation method for equivalent relative dielectric constant (ERDC) based on the model test results and validated its rationality through the analysis of numerical simulation results and real GPR waveforms.

Transmission line fault cause identification method based on transient waveform image and MCNN-LSTM

Power system operators commonly employ fault recorders to record and analyze transmission line faults. However, the length of transmission lines, the number and location of power sources and the sampling frequency of fault recorder will influence the characteristics of the fault waveform leading to fault identification error. This paper proposed a multi-head convolutional neural network with long short-term memory (MCNN-LSTM) using transient waveform image as input for transmission line fault type and fault cause classification. First, the characteristics of transmission line faults were analyzed. Second, MCNN-LSTM is designed by considering interaction between electrical signal. Finally, the modified IEEE-39 node power network was adopted to test the proposed method. The test results show that: this method can improve the identification accuracy of transmission line faults even if there is class unbalance in fault sample set, and shows strong adaptability when transferred to small-size fault samples of transmission lines.

Machine learning discriminates P2X7-mediated intracellular calcium sparks in human-induced pluripotent stem cell-derived neural stem cells

Adenosine triphosphate (ATP) is an extracellular signaling molecule that mainly affects the pathophysiological situation in the body and can be sensed by purinergic receptors, including ionotropic P2X7. Neuronal stem cells (NSCs) remain in adult neuronal tissues and can contribute to physiological processes via activation by evoked pathophysiological situations. In this study, we revealed that human-induced pluripotent stem cell-derived NSCs (iNSCs) have ATP-sensing ability primarily via the purinergic and ionotropic receptor P2X7. Next, to develop a machine learning (ML)-based screening system for food-derived neuronal effective substances and their effective doses, we collected ATP-triggered calcium responses of iNSCs pretreated with several substances and doses. Finally, we discovered that ML was performed using composite images, each containing nine waveform images, to achieve a better ML model (MLM) with higher precision. Our MLM can correctly sort subtle unidentified changes in waveforms produced by pretreated iNSCs with each substance and/or dose into the positive group, with common mRNA expression changes belonging to the gene ontology signatures.

Partial Discharge Pattern Recognition Method Based on Transfer Learning and DenseNet Model

With the development of intelligent sensing technology, a large amount of partial discharge (PD) time-domain waveform images are generated in the on-site detection of gas insulated switchgear (GIS) partial discharge. Traditional pattern recognition methods are mostly aimed at structured data and cannot directly identify defect types of such data.At the same time, the deep learning method for GIS PD pattern recognition is generally faced with the problem of small samples.In order to solve the above problems, this paper proposes a PD pattern recognition method based on transfer learning and DenseNet model.Firstly, the time-domain waveform images are processed by image enhancement, normalization, image compression and other image processing techniques.The FDTD method was used to simulate GIS PD, and the time domain waveform image database of four PD defects are established.Using CNN and transfer learning, the recognition accuracy of the model is increased to 95%, with better robustness.The recognition performance of different CNN structures is studied. The results show that DenseNet model has higher accuracy than other structures and shorter training time. This study can be used to diagnose the insulation status of GIS equipment in-site.

Tutorial: Real-time coherent terahertz imaging of objects moving in one direction with constant speed

Pulsed terahertz (THz) electric fields enable various coherent THz imaging modes, such as reflection tomography, spectral imaging, and computed tomography (CT) for nondestructive inspection, quality control, and material characterization. The extension of coherent THz imaging modes to moving objects has been regarded as key to their social implementation. This Tutorial focuses on two-dimensional spatiotemporal (2D-ST) THz imaging of objects moving in one direction with constant speed as a promising means of enabling real-time coherent THz imaging. In 2D-ST THz imaging, the temporal waveform and line image of the THz pulse are simultaneously acquired without the need for mechanical scanning of the time delay and sample position using a combination of non-collinear 2D free-space electro-optic sampling with THz line-imaging optics. This 2D-ST THz imaging boosts the imaging rates of THz reflection tomography, THz spectral imaging, and THz CT to levels that are applicable to moving objects. The advanced THz reflection tomography and THz spectral imaging that result from the assistance of 2D-ST THz imaging achieve real-time line imaging of cross sections and spectral signatures, respectively. Subsequently, this enables in-line total inspection of objects moving on a translation stage or a conveyor belt. A THz CT system using real-time line projection of a THz beam is effectively applied to a 2D spectral cross section of a continuously rotating object. 2D-ST THz imaging enables the functional THz imaging of moving objects in various practical applications.

Ultrasound computed tomography based on full waveform inversion with source directivity calibration

Ultrasound computed tomography based on full waveform inversion has the potential to provide high-resolution images of human tissues in a quantitative manner. A successful ultrasound computed tomography system requires the decent knowledge of acquisition array, including the spatial position and the directivity of each transducer, to meet the high-level demand of clinical applications. The conventional full waveform inversion algorithm assumes a point source with the omni-directional emission. Such assumption does not hold when the directivity of emitting transducer is not negligible. For a practical implementation, an efficient and accurate self-checking evaluation of directivity is crucial prior to the reconstruction of images. We propose to measure the directivity of each emitting transducer using the full-matrix captured data obtained with a water-immersed and target-free experiment. We introduce the weighted virtual point-source array to act as the proxy of emitting transducer during the numerical simulation. The weights of different points in the virtual array can be calculated from the observed data using the gradient-based local optimization method. Although the full waveform imaging method relies on the finite-difference solver of wave equation, such directivity estimation benefits from the introduction of analytical solver. The trick significantly reduces the numerical cost, enabling an automatic directivity self-check at boot. We verify the feasibility, efficiency, and accuracy of the virtual array method through simulated and experimental tests. For the experimental test, we also illustrate that full waveform inversion with directivity calibration can reduce the artifacts introduced by the conventional point source assumption, improving the quality of reconstructed images..

A Multi-Channel Ensemble Method for Error-Related Potential Classification Using 2D EEG Images

An error-related potential (ErrP) occurs when people's expectations are not consistent with the actual outcome. Accurately detecting ErrP when a human interacts with a BCI is the key to improving these BCI systems. In this paper, we propose a multi-channel method for error-related potential detection using a 2D convolutional neural network. Multiple channel classifiers are integrated to make final decisions. Specifically, every 1D EEG signal from the anterior cingulate cortex (ACC) is transformed into a 2D waveform image; then, a model named attention-based convolutional neural network (AT-CNN) is proposed to classify it. In addition, we propose a multi-channel ensemble approach to effectively integrate the decisions of each channel classifier. Our proposed ensemble approach can learn the nonlinear relationship between each channel and the label, which obtains 5.27% higher accuracy than the majority voting ensemble approach. We conduct a new experiment and validate our proposed method on a Monitoring Error-Related Potential dataset and our dataset. With the method proposed in this paper, the accuracy, sensitivity and specificity were 86.46%, 72.46% and 90.17%, respectively. The result shows that the AT-CNNs-2D proposed in this paper can effectively improve the accuracy of ErrP classification, and provides new ideas for the study of classification of ErrP brain-computer interfaces.