Input To Neural Network Research Articles

Predicting missing sonic logs with seismic constraint

Compressional and shear sonic transit time logs (DT and DTS, respectively) provide important petrophysical and geomechanical information for subsurface characterization. However, they are often not acquired in all wells because of cost limitations or borehole problems. We develop a method to estimate the DT and DTS simultaneously from other commonly acquired well logs such as gamma ray, density, and neutron porosity. Our method consists of two consecutive models to predict the sonic logs and the seismic traces at the well locations. The model predicting the seismic traces adds a spatial constraint to the model predicting the sonic logs. Our method also quantifies the uncertainties of the prediction, which come from the uncertainties of neural network parameters and input data. We train the network on four wells from the Poseidon data set located on the Australian shelf in the Browse Basin. We test the network on the other two wells from the Browse Basin. The test results indicate better predictions of sonic logs when we add the seismic constraint.

Adversarial Camera Patch: An Effective and Robust Physical-World Attack on Object Detectors

Physical adversarial attacks present a novel and growing challenge in cybersecurity, especially for systems reliant on physical inputs for Deep Neural Networks (DNNs), such as those found in Internet of Things (IoT) devices. They are vulnerable to physical adversarial attacks where real-world objects or environments are manipulated to mislead DNNs, thereby threatening the operational integrity and security of IoT devices. The camera-based attacks are one of the most practical adversarial attacks, which are easy to implement and more robust than all the other attack methods, and pose a big threat to the security of IoT. This paper proposes Adversarial Camera Patch (ADCP), a novel approach that employs a single-camera patch to launch robust physical adversarial attacks against object detectors. ADCP optimizes the physical parameters of the camera patch using Particle Swarm Optimization (PSO) to identify the most adversarial configuration. The optimized camera patch is then attached to the lens to generate stealthy and robust adversarial samples physically. The effectiveness of the proposed approach is validated through ablation experiments in a digital environment, with experimental results demonstrating its effectiveness even under worst-case scenarios (minimal width, maximum transparency). Notably, ADCP exhibits higher robustness in both digital and physical domains compared to the baseline. Given the simplicity, robustness, and stealthiness of ADCP, we advocate for attention towards the ADCP framework as it offers a means to achieve streamlined, robust, and stealthy physical attacks. Our adversarial attacks pose new challenges and requirements for cybersecurity.

Gas concentration prediction by LSTM network combined with wavelet thresholding denoising and phase space reconstruction

The Long Short-Term Memory neural network is a specialized architecture designed for handling time series data, extensively applied in the field of predicting gas concentrations. In the harsh conditions prevalent in coal mines, the time series data of gas concentrations collected by sensors are susceptible to noise interference. Directly inputting such noisy data into a neural network for training would significantly reduce predictive accuracy and lead to deviations from the actual values. The Empirical Mode Decomposition method, commonly employed in gas concentration prediction, faces challenges in practical engineering applications due to the substantial influence of newly acquired data on the initial decomposition subsequence values. Consequently, it is difficult to use this method as intended. Conversely, the Wavelet Threshold Denoising method does not encounter this issue. Furthermore, gas concentration sequences exhibit chaotic characteristics. Performing phase space reconstruction allows for the extraction of additional valuable hidden information. In light of these factors, a prediction model is proposed, integrating WTD, Phase Space Reconstruction, and LSTM neural networks. Initially, the gas concentration sequence itself is subjected to wavelet threshold denoising. Subsequently, phase space reconstruction is performed, and the resulting reconstructed phase space matrix serves as the input for the LSTM neural network. The outcomes from the final LSTM neural network reveal that the PS method indeed extracts more valuable information. The Mean Absolute Error and Root Mean Square Error are reduced by 35.1% and 25%, respectively. Additionally, when compared to the PS-LSTM model without utilizing the WTD method, the WTD-PS-LSTM predictive model showcases reductions of 77.1% and 80% in MAE and RMSE, respectively. Compared with the LSTM model, the MAE and RMSE of the WTD-PS-LSTM prediction model were reduced by 81.4% and 82.6%, respectively. This greatly improves the credibility of whether or not a response related to coal mine safety management is implemented.

Artificial Intelligence-Assisted RFID Tag-Integrated Multi-Sensor for Quality Assessment and Sensing.

Radio frequency identification (RFID) is well known as an identification, track, and trace approach and is considered to be the key physical layer technology for the industrial internet of things (IIoT). However, IIoT systems have to introduce additional complex sensor networks for pervasive monitoring, and there are still challenges related to item-level sensing and data recording. To overcome the shortage, this work proposes an artificial intelligence (AI)-assisted RFID-based multi-sensing technology. Both passive and semi-passive RFID tag-integrated multi-sensors are developed. The main contributions and the novelty of this investigation are as follows. A UHF RFID tag-integrated multi-sensor with a boosted charge pump is proposed; it enables high RF signal sensitivity and a long operational range. The whole hardware design, including the antenna and energy harvester, are studied. Moreover, a demonstration with real-world ham product sensing data is conducted. This work also proposes and successfully demonstrates the integration of machine learning algorithms, specifically the NARX neural network, with RFID sensing data for food product quality assessment and sensing (QAS). This application of machine learning to RFID-generated data for quality assessment is also a novel aspect of the research. The deployment of an autoregressive model with an exogenous input (NARX) neural network model, tailored for nonlinear processes, emerges as the most effective, achieving a root mean square error (RMSE) of 0.007 and an R-squared value of 0.99 for ham product QAS. By deploying the technology, low-cost, timely, and flexible product QAS can be achieved in manufacturing industries, which helps product quality improvement and the optimization of the manufacturing line and supply chain.

Forecasting urban air pollution using multi-site spatiotemporal data fusion method (Geo-BiLSTMA)

The issue of air pollution is receiving more and more attention. And the key to effective monitoring and prediction of air pollution lies in combining spatial and temporal information from monitoring site data. In this study, we integrating the spatiotemporal characteristics of several sparse and irregular monitoring stations, which is essential for getting accurate air pollution prediction. Krigin interpolation method is used to generate geographical uniform grid surface data, then the historical data of the target site and its surrounding eight neighbors are extracted as the input of BiLSTM neural network which inducing attention mechanism, so we named it Geo-BiLSTMA. We experimented with the algorithm on air pollution datasets in Xi'an, and the results of the algorithm were compared with the five baseline algorithms. We also verified that the Geo-BiLSTMA model has a good generalization ability during the winter and summer prevention periods. And the algorithm has been deployed as an auxiliary model in the task of the national key R&D project.

Evaluation of Neural Network Effectiveness on Sliding Mode Control of Delta Robot for Trajectory Tracking

The Delta robot is an over-actuated parallel robot with highly nonlinear kinematics and dynamics. Designing the control for a Delta robot to carry out various operations is a challenging task. Various advanced control algorithms, such as adaptive control, sliding mode control, and model predictive control, have been investigated for trajectory tracking of the Delta robot. However, these control algorithms require a reliable input–output model of the Delta robot. To address this issue, we have created a control-affine neural network model of the Delta robot with stepper motors. This is a completely data-driven model intended for control design consideration and is not derivable from Newton’s law or Lagrange’s equation. The neural networks are trained with randomly sampled data in a sufficiently large workspace. The sliding mode control for trajectory tracking is then designed with the help of the neural network model. Extensive numerical results are obtained to show that the neural network model together with the sliding mode control exhibits outstanding performance, achieving a trajectory tracking error below 5 cm on average for the Delta robot. Future work will include experimental validation of the proposed neural network input–output model for control design for the Delta robot. Furthermore, transfer learnings can be conducted to further refine the neural network input–output model and the sliding mode control when new experimental data become available.

Data-driven modeling for residual velocity of projectile penetrating reinforced concrete slabs

The penetration of projectiles into reinforced concrete (RC) structures is crucial for defense technology. However, accurately estimating the resistance of RC structures to projectile penetration is challenging due to complex parameter nonlinearities. This paper proposes a data-driven modeling methodology that combines validated finite element (FE) simulations with machine learning (ML) algorithms to investigate the penetration resistance of RC slabs against projectiles. The methodology employs an extreme learning machine (ELM) network to identify extreme conditions characterizing penetration resistance, reducing data requirements, and improving prediction efficiency. Predicted results from the ELM network are then used as input for a recurrent neural network (RNN) to predict residual velocity curves over time, a crucial metric for evaluating penetration resistance. To cover various loading scenarios, 64 sets of FE models with different incident velocities, caliber-radius-head (CRH), and projectile masses are developed based on available experimental studies. The Johnson-Holmquist II model and Johnson-Cook model are employed for concrete and steel, respectively, to describe material behavior at high strain rates. The FE prediction results form a training sample database, and ML models are established using CRH, incident velocity, and projectile mass as input layers and projectile residual velocity as the output layer. The results demonstrate high Pearson correlation coefficients of 0.984 and 0.991 for the ELM and RNN prediction results, respectively, with numerical simulation results. The proposed networks provide accurate and efficient alternatives to time-consuming FE simulations for predicting projectile penetration performance. This strategy offers a new ML-based method for rapid evaluation, particularly for in-situ applications.

Theory-driven neural network for total organic carbon content estimation from well logs

Abstract Total organic carbon content (TOC) is a crucial geochemical indicator for assessing the hydrocarbon generation potential of source rock formations. Traditional TOC evaluation methods employing well logs do not match well with measured data in complex formations. While intelligent algorithms have the potential to improve TOC estimation accuracy, they have poor petrophysical significance. In this study, we develop the theory-driven neural networks (TdNN) to extend TOC prediction accuracy via combining logging mechanism models into the input of neural network. TdNN have two ways for merging petrophysical models with intelligent model, specifically TdNN based on series connection (STdNN) and based on embedded way (ETdNN). STdNN processes log responses through logging mechanism models and subsequently inputs them to the neural network for TOC evaluation. ETdNN inputs log responses and calculation results from logging mechanism models to the neural network to estimate TOC. Two ways achieve more accurate prediction results than ordinary neural network via improving the quality of input data, analogous to applying data augmentation methods on intelligent models. This study utilizes the Schmoker method, 𝛿logR method, and multivariate regression method as logging mechanism models of TdNN. The prediction performance of the proposed models is tested using data from two shale reservoirs in the Longmaxi Formation and the Shahejie Formation. Results show that the proposed ETdNN achieves the highest prediction accuracy, meaning that the ETdNN is suitable for predicting TOC of shale formations.

Subtracting-adding strategy for necrotic lesion segmentation in osteonecrosis of the femoral head.

Osteonecrosis of the femoral head (ONFH) is a severe bone disease that can progressively lead to hip dysfunction. Accurately segmenting the necrotic lesion helps in diagnosing and treating ONFH. This paper aims at enhancing deep learning models for necrosis segmentation. Necrotic lesions of ONFH are confined to the femoral head. Considering this domain knowledge, we introduce a preprocessing procedure, termed the "subtracting-adding" strategy, which explicitly incorporates this domain knowledge into the downstream deep neural network input. This strategy first removes the voxels outside the predefined volume of interest to "subtract" irrelevant information, and then it concatenates the bone mask with raw data to "add" anatomical structure information. Each of the tested off-the-shelf networks performed better with the help of the "subtracting-adding" strategy. The dice similarity coefficients increased by 10.93%, 9.23%, 9.38% and 1.60% for FCN, HRNet, SegNet and UNet, respectively. The improvements in FCN and HRNet were statistically significant. The "subtracting-adding" strategy enhances the performance of general-purpose networks in necrotic lesion segmentation. This strategy is compatible with various semantic segmentation networks, alleviating the need to design task-specific models.

Graph Autoencoders for Embedding Learning in Brain Networks and Major Depressive Disorder Identification.

Brain functional connectivity (FC) networks inferred from functional magnetic resonance imaging (fMRI) have shown altered or aberrant brain functional connectome in various neuropsychiatric disorders. Recent application of deep neural networks to connectome-based classification mostly relies on traditional convolutional neural networks (CNNs) using input FCs on a regular Euclidean grid to learn spatial maps of brain networks neglecting the topological information of the brain networks, leading to potentially sub-optimal performance in brain disorder identification. We propose a novel graph deep learning framework that leverages non-Euclidean information inherent in the graph structure for classifying brain networks in major depressive disorder (MDD). We introduce a novel graph autoencoder (GAE) architecture, built upon graph convolutional networks (GCNs), to embed the topological structure and node content of large fMRI networks into low-dimensional representations. For constructing the brain networks, we employ the Ledoit-Wolf (LDW) shrinkage method to efficiently estimate high-dimensional FC metrics from fMRI data. We explore both supervised and unsupervised techniques for graph embedding learning. The resulting embeddings serve as feature inputs for a deep fully-connected neural network (FCNN) to distinguish MDD from healthy controls (HCs). Evaluating our model on resting-state fMRI MDD dataset, we observe that the GAE-FCNN outperforms several state-of-the-art methods for brain connectome classification, achieving the highest accuracy when using LDW-FC edges as node features. The graph embeddings of fMRI FC networks also reveal significant group differences between MDD and HCs. Our framework demonstrates the feasibility of learning graph embeddings from brain networks, providing valuable discriminative information for diagnosing brain disorders.

Research on multiple calibration methods for measuring error of indicator value of withstand voltage tester

Voltage resistance testing often needs to be carried out under specific environmental conditions, such as high pressure and high temperature. These complex environments may interfere with the test equipment and calibration process, thus affecting the accuracy of calibration results. Therefore, it is necessary to ensure that the tester can adapt to various environmental conditions and make calibration adjustments accordingly. Therefore, a multi-calibration method based on the BP neural network is proposed to calibrate the error of the indicator value of the voltage tester. The power input and output characteristic function of the withstand voltage tester is established, and the output of each layer is described as the input of the BP neural network. Based on the adjustment results, multiple calibration steps for the measuring error of the voltage tester are designed. The experimental results show that the method can calibrate the measuring error of the voltage tester at different positions at the same time, and the application effect is good.

Physics-informed neural networks for transonic flow around a cylinder with high Reynolds number

The physics-informed neural network (PINN) method is extended to learn and predict compressible steady-state aerodynamic flows with a high Reynolds number. To better learn the thin boundary layer, the sampling distance function and hard boundary condition are explicitly introduced into the input and output layers of the deep neural network, respectively. A gradient weight factor is considered in the loss function to implement the PINN methods based on the Reynolds averaged Navier–Stokes (RANS) and Euler equations, respectively, denoted as PINN–RANS and PINN–Euler. Taking a transonic flow around a cylinder as an example, these PINN methods are first verified for the ability to learn complex flows and then are applied to predict the global flow based on a part of physical data. When predicting the global flow based on velocity data in local key regions, the PINN–RANS method can always accurately predict the global flow field including the boundary layer and wake, while the PINN–Euler method can accurately predict the inviscid region. When predicting the subsonic and transonic flows under different freestream Mach numbers (Ma∞= 0.3–0.7), the flow fields predicted by both methods avoid the inconsistency with the real physical phenomena of the pure data-driven method. The PINN–RANS method is insufficient in shock identification capabilities. Since the PINN–Euler method does not need the second derivative, the training time of PINN–Euler is only 1/3 times that of PINN–RANS at the same sampling point and deep neural network.

A physics-guided machine learning framework for real-time dynamic wake prediction of wind turbines

Efficient and accurate prediction of the wind turbine dynamic wake is crucial for active wake control and load assessment in wind farms. This paper proposes a real-time dynamic wake prediction model for wind turbines based on a physics-guided neural network. The model can predict the instantaneous dynamic wake field under various operating conditions using only the inflow wind speed as input. The model utilizes Taylor's frozen-flow hypothesis and a steady-state wake model to convert instantaneous inflow wind speed and turbine parameters into neural network input features. A deep convolutional neural network then maps these features to desired wake field snapshots, enabling dynamic wake predictions for wind turbines. To train the model, we generated approximately 255 000 instantaneous flow field snapshots of single-turbine wakes using the large eddy simulation, covering different thrust coefficients and yaw angles. The model was trained using the supervised learning method and verified on the test set. The results indicate that the model can effectively predict the dynamic wake characteristics, including the dynamic wake meandering and the wake deflection of the yawed turbines. The model can also assess both the instantaneous wake velocity and the instantaneous wake center of a wind turbine. At a thrust coefficient of 0.75, the root mean square error for the predicted instantaneous wake velocity is around 6.53%, while the Pearson correlation coefficient for the predicted instantaneous wake center can reach 0.624. Furthermore, once the model is trained, its prediction accuracy does not decrease with the increase in the time span.

Gearbox fault diagnosis method based on improved semi-supervised MTDL and GAF

Aiming at the problem that it takes a long time and high cost to obtain complete labeled data under intelligent fault diagnosis and unlabeled data is not used. This paper proposes an improved semi-supervised mean teacher deep learning (MTDL) and Gramian angle field (GAF) fusion diagnostic method. This method fully utilizes a small number of labeled samples and a large number of unlabeled samples to deeply mine invisible fault features and potential physical correlations. At the same time, it solves the problem of losing the inter-data correlation structure when one-dimensional time series signals are used as inputs for neural networks. The GAF-MTDL method uses consistency regularization and modifies the network structure in the mean teacher algorithm into a semi-supervised deep learning model enhanced by WideResNet. The experimental results show that the proposed GAF-MTDL method saves a lot of manual labeling costs, improves the recognition accuracy and generalization ability, and can achieve excellent prediction accuracy with very little labeled data. In the end, the accuracy of planetary gear fault identification reached 98.22% under the labeling rate of 20%, and the accuracy of fault identification reached 99.98% through the verification of the bearing data set of Case Western Reserve University. The value of this research is to bring an efficient and low-cost technology to the field of industrial intelligent fault diagnosis, which can significantly improve the accuracy of fault identification.

Shared input and recurrency in neural networks for metabolically efficient information transmission.

Shared input to a population of neurons induces noise correlations, which can decrease the information carried by a population activity. Inhibitory feedback in recurrent neural networks can reduce the noise correlations and thus increase the information carried by the population activity. However, the activity of inhibitory neurons is costly. This inhibitory feedback decreases the gain of the population. Thus, depolarization of its neurons requires stronger excitatory synaptic input, which is associated with higher ATP consumption. Given that the goal of neural populations is to transmit as much information as possible at minimal metabolic costs, it is unclear whether the increased information transmission reliability provided by inhibitory feedback compensates for the additional costs. We analyze this problem in a network of leaky integrate-and-fire neurons receiving correlated input. By maximizing mutual information with metabolic cost constraints, we show that there is an optimal strength of recurrent connections in the network, which maximizes the value of mutual information-per-cost. For higher values of input correlation, the mutual information-per-cost is higher for recurrent networks with inhibitory feedback compared to feedforward networks without any inhibitory neurons. Our results, therefore, show that the optimal synaptic strength of a recurrent network can be inferred from metabolically efficient coding arguments and that decorrelation of the input by inhibitory feedback compensates for the associated increased metabolic costs.

Electromyography-Based Biomechanical Cybernetic Control of a Robotic Fish Avatar

This study introduces a cybernetic control and architectural framework for a robotic fish avatar operated by a human. The behavior of the robot fish is influenced by the electromyographic (EMG) signals of the human operator, triggered by stimuli from the surrounding objects and scenery. A deep artificial neural network (ANN) with perceptrons classifies the EMG signals, discerning the type of muscular stimuli generated. The research unveils a fuzzy-based oscillation pattern generator (OPG) designed to emulate functions akin to a neural central pattern generator, producing coordinated fish undulations. The OPG generates swimming behavior as an oscillation function, decoupled into coordinated step signals, right and left, for a dual electromagnetic oscillator in the fish propulsion system. Furthermore, the research presents an underactuated biorobotic mechanism of the subcarangiform type comprising a two-solenoid electromagnetic oscillator, an antagonistic musculoskeletal elastic system of tendons, and a multi-link caudal spine composed of helical springs. The biomechanics dynamic model and control for swimming, as well as the ballasting system for submersion and buoyancy, are deduced. This study highlights the utilization of EMG measurements encompassing sampling time and μ-volt signals for both hands and all fingers. The subsequent feature extraction resulted in three types of statistical patterns, namely, Ω,γ,λ, serving as inputs for a multilayer feedforward neural network of perceptrons. The experimental findings quantified controlled movements, specifically caudal fin undulations during forward, right, and left turns, with a particular emphasis on the dynamics of caudal fin undulations of a robot prototype.

Abstract 65: Automated Detection of Facial Droop: Ml/AI Based Tool for Early Detection of Stroke

Introduction: Facial droop (FD) is a common symptom (45-60%) of stroke so early detection is critical for timely treatment. FD is a part of almost all stroke scales. Our study is to develop a machine learning (ML)/AI based tool that uses an automated detection of FD as one of the components. Our approach uses a normalized value of minor facial expressions called a blendshape. Methods: Google’s Mediapipe was used to create 52 different blendshape values (e.g., left eyebrow raised, right mouth lowered, etc.) from an input image, which were then used as inputs for a simple deep neural network (DNN). We trained this DNN using the blendshapes from 872 images of people with symptoms of facial palsy (FP) and 872 images of people without symptoms exhibiting multiple emotions (neutral, smiling, frowning, etc.). Results: The proposed approach was evaluated on a dataset of facial images of patients with FD and strokes. Our model showed promising results with over 91% test accuracy. This result suggests that the proposed approach is a promising new method for the early detection of FP with high accuracy. The approach is integrated in our ML/AI based tool to replicate different stroke scales. Conclusions: Our model shows that the use of Blendshapes as features in a ML classifier can produce great results in the detection of FP. For greater accuracy, a better dataset with higher quality images as well as unilateral facial expressions (winking, smirks, etc.) should be used. In a future study, a ML model that uses Blendshapes as inputs could be used to classify FP on a scale.

Prediction of ionospheric total electron content data using NARX neural network model

Successful prediction of ionospheric total electron content (TEC) data will help in correction of positioning errors in global navigation satellite systems (GNSS) caused by the ionosphere. This research paper proposes a prediction model for ionospheric TEC using a nonlinear autoregressive with exogenous inputs (NARX) neural network that utilizes past TEC data alongwith solar and geomagnetic indices namely F10.7, disturbed storm (Dst), Kp, Ap, and time of the day. We assess the prediction capability of our model at different latitudes during different solar activity years. We compare our results with another NARX model which uses previous TEC data along with time of the day, day of the year and season as exogenous parameters. The results show that for the solar minimum year the TEC prediction accuracy improves by 35.71% and for the solar maximum year it improves by 31.20%. The results using root mean square error (RMSE), mean absolute error (MAE), correlation coefficient, and symmetric mean absolute percentage error (sMAPE) clearly indicate that solar and geomagnetic indices along with time of the day help in enhancing prediction accuracy of TEC across different latitudinal regions during both solar minimum and maximum years.

Dynamic load prediction of charging piles for energy storage electric vehicles based on Space-time constraints in the internet of things environment

Abstract This paper puts forward the dynamic load prediction of charging piles of energy storage electric vehicles based on time and space constraints in the Internet of Things environment, which can improve the load prediction effect of charging piles of electric vehicles and solve the problems of difficult power grid control and low power quality caused by the randomness of charging loads in time and space. After constructing a traffic road network model based on the Internet of Things, a travel chain model with different complexity and an electric vehicle charging model, the travel chain is randomly extracted. With the shortest travel time as a constraint, combined with the traffic road network model based on the Internet of Things, the travel route and travel time are determined. According to the State of Charge (SOC) and the travel destination, the location and charging time of the energy storage electric vehicle charging pile are determined. After obtaining the time-space distribution information of the energy storage electric vehicle charging pile at different times and in different regions, it is used as the input of the deep multi-step time-space dynamic neural network, and the network output is the dynamic electric vehicle charging pile. The experimental results show that this method can realize the dynamic load prediction of electric vehicle charging piles. When the number of stacking units is 11, the indexes of Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) are the lowest and the index of R 2 is the largest. The load of charging piles in residential areas and work areas exists in the morning and evening peak hours, while the load fluctuation of charging piles in other areas presents a decentralized change law; The higher the complexity of regional traffic network, the greater the load of electric vehicle charging piles in the morning rush hour.

Research on modulation recognition method of electromagnetic signal based on wavelet transform convolutional neural network

The method of electromagnetic signal modulation recognition based on wavelet transform convolutional neural network is studied to improve the effect of electromagnetic signal modulation recognition. By analyzing the electromagnetic signal modulation model, the original electromagnetic signal is preprocessed by wavelet transform to remove the noise of the original electromagnetic signal. The processed electromagnetic signal is used as the input of convolutional neural network, and the electromagnetic signal feature vector is extracted through the convolution layer of convolutional neural network. By using full connection operation, the advanced feature vector of electromagnetic signal is integrated, and the electromagnetic signal is classified by softmax function, and the electromagnetic signal modulation recognition result is output, thus realizing the electromagnetic signal modulation recognition. The experimental results show that when the number of layers of wavelet decomposition is 7 and the wavelet function is Db9, the wavelet transform has the best denoising effect on electromagnetic signal data. At the same time, the network training efficiency of this method is high, and the accuracy of electromagnetic signal modulation recognition is as high as 97.2 %, which improves the effect of electromagnetic signal modulation recognition and is suitable for various types of electromagnetic signal modulation recognition.