Trained Support Vector Machine Model Research Articles

Tri-axial accelerometry allows to determine parental food provisioning behaviour in a marine bird

The study of parental food provisioning is essential for understanding the breeding ecology of birds. We conducted the first study using accelerometry to detect food provisioning in birds, using Support Vector Machine (SVM) models to identify when adults feed chicks of three different age classes. Accelerometers were attached to the head of adult female Imperial Shags (Leucocarbo atriceps), and various attributes derived from the acceleration signals were used to train SVM models for each chick age class. Model performance improved with chick age class, with SVM models achieving high overall accuracy (>88%) and highest sensitivity in older chick categories (>91%). However, precision values, especially for younger chicks, remained relatively low (between 26% and 45%). The application of a time filter based on the minimum duration of the observed food provisioning behaviours for each chick age category, improved model performance by reducing false provisioning behaviours, particularly in the model for older chicks, which showed the highest precision (72.4%). This study highlights the effectiveness of accelerometry and machine learning in studying parental food provisioning in birds, providing a rapid and accurate data collection method to complement traditional techniques. The described methodology can be applied to any bird species that exhibits distinctive movements while feeding its offspring and has suitable characteristics for attaching an accelerometer to the body part that best captures this movement. Finally, it is hoped that the results of this study will contribute to future research on key questions in parental investment theory and reproductive strategies in birds.

SVM-assisted ANN model with principal component analysis based dimensionality reduction for enhancing state-of-charge estimation in LiFePO4 batteries

Accurate estimation of state-of-charge (SoC) is significant for monitoring the operation of LiFePO4 batteries. This paper addresses the challenges posed by the nonlinear characteristics of LiFePO4 batteries, which adversely affect the accuracy of SoC estimation. The proposed novel methodology for SoC estimation in LiFePO4 batteries involves a Support Vector Machine (SVM) assisted Artificial Neural Network (ANN) model incorporating principal component analysis (PCA) to efficiently interpret the input data. To improve the prediction accuracy, the non-linear characteristics of LiFePO4 are divided into three parts and each of these parts is used to train a separate ANN model. Further, an optimal number of hidden layers and neurons are selected for each ANN model to minimize the prediction errors. The usage of dedicated smaller datasets for each ANN model simplifies the structure. SVM classifies the battery operating regions and selects the most suitable ANN model for SoC estimation. PCA is applied to process the obtained experimental data resulting in three principal components serving as inputs for the SVM-assisted ANN model. Further, with PCA the input dimensions are reduced from four to three, thereby leading to computational simplicity. The input data comprises of current, voltage, open-circuit voltage and temperature of the battery. An experimental prototype, comprising a customized battery pack and sensing mechanisms, is developed for data collection for training SVM and ANN models. With the proposed SVM-assisted ANN involving PCA, the loss function is minimized and an average Root Mean Square Error (RMSE) of 0.3133 is achieved. This demonstrates the feasibility, accuracy and applicability of SoC estimation with the developed SVM-assisted ANN model for LiFePO4 batteries.

The performance of Logistic Regression, Decision Tree, KNN, Naive Bayes and SVM for identifying Automotive Cybersecurity Attack and Prevention: An Experimental Study

The automotive industry has witnessed a significant increase in cyber threats as vehicles become more connected and reliant on software-driven systems. To safeguard against these threats, effective cybersecurity measures must be implemented. This paper explores the use of Support Vector Machine (SVM) algorithms as a means of bolstering automotive cybersecurity attack prevention. SVM algorithms have demonstrated remarkable capabilities in various domains due to their ability to handle complex and high-dimensional data. By leveraging SVM algorithms, this research aims to enhance the detection and prevention of cybersecurity attacks targeting automotive systems.
 The proposed approach involves training SVM models using labeled datasets that include both normal and anomalous driving scenarios. By analyzing various features and patterns extracted from the data, the SVM models can learn to differentiate between normal and malicious behaviour. These models can then be used in real-time to identify and mitigate potential cyber threats.
 The results of this study is compared with other machine learning algorithm like Logistic Regression, Naive Bayes, KNN, Decision Tree, highlight the effectiveness of SVM algorithms in detecting and preventing automotive cybersecurity attacks. The models exhibit high accuracy rates in distinguishing between normal and anomalous behavior, providing a robust defense mechanism against potential threats.
 In conclusion, this research emphasizes the significance of leveraging SVM learning algorithms for automotive cybersecurity attack prevention. By harnessing the power of machine learning, automotive systems can be fortified against cyber threats, ensuring the safety and integrity of vehicles and their passengers.

Sensor-Based Quantification of MDS-UPDRS III Subitems in Parkinson's Disease Using Machine Learning.

Wearable sensors could be beneficial for the continuous quantification of upper limb motor symptoms in people with Parkinson's disease (PD). This work evaluates the use of two inertial measurement units combined with supervised machine learning models to classify and predict a subset of MDS-UPDRS III subitems in PD. We attached the two compact wearable sensors on the dorsal part of each hand of 33 people with PD and 12 controls. Each participant performed six clinical movement tasks in parallel with an assessment of the MDS-UPDRS III. Random forest (RF) models were trained on the sensor data and motor scores. An overall accuracy of 94% was achieved in classifying the movement tasks. When employed for classifying the motor scores, the averaged area under the receiver operating characteristic values ranged from 68% to 92%. Motor scores were additionally predicted using an RF regression model. In a comparative analysis, trained support vector machine models outperformed the RF models for specific tasks. Furthermore, our results surpass the literature in certain cases. The methods developed in this work serve as a base for future studies, where home-based assessments of pharmacological effects on motor function could complement regular clinical assessments.

Deep Texture Analysis—Enhancing CT Radiomics Features for Prediction of Head and Neck Cancer Treatment Outcomes: A Machine Learning Approach

(1) Background: Some cancer patients do not experience tumour shrinkage but are still at risk of experiencing unwanted treatment side effects. Radiomics refers to mining biomedical images to quantify textural characterization. When radiomics features are labelled with treatment response, retrospectively, they can train predictive machine learning (ML) models. (2) Methods: Radiomics features were determined from lymph node (LN) segmentations from treatment-planning CT scans of head and neck (H&N) cancer patients. Binary treatment outcomes (complete response versus partial or no response) and radiomics features for n = 71 patients were used to train support vector machine (SVM) and k-nearest neighbour (k-NN) classifier models with 1–7 features. A deep texture analysis (DTA) methodology was proposed and evaluated for second- and third-layer radiomics features, and models were evaluated based on common metrics (sensitivity (%Sn), specificity (%Sp), accuracy (%Acc), precision (%Prec), and balanced accuracy (%Bal Acc)). (3) Results: Models created with both classifiers were found to be able to predict treatment response, and the results suggest that the inclusion of deeper layer features enhanced model performance. The best model was a seven-feature multivariable k-NN model trained using features from three layers deep of texture features with %Sn = 74%, %Sp = 68%, %Acc = 72%, %Prec = 81%, %Bal Acc = 71% and with an area under the curve (AUC) the receiver operating characteristic (ROC) of 0.700. (4) Conclusions: H&N Cancer patient treatment-planning CT scans and LN segmentations contain phenotypic information regarding treatment response, and the proposed DTA methodology can improve model performance by enhancing feature sets and is worth consideration in future radiomics studies.

Particle swarm optimization support vector machine-based coal and rock cutting tool load spectrum identification method

The goal of this research is to achieve safe and efficient excavation of coal and rock tunnels with complex geological structures, and to enhance the self-sensing ability of coal and rock cutting equipment and tools. Particle swarm optimization support vector machine is used to identify the cutting state of disc cutting tools. EDEM finite element analysis software is used to analyze cutting process characteristics of the disc cutting tool when used to cut through coal and rock with different compressive strengths. Empirical mode decomposition is used to decompose the load spectrum characteristics; for this purpose, the first-order and seventh-order intrinsic mode functions containing all the feature information of the original signal of the load spectrum are selected. The sample entropy is calculated as the feature input vector. The extracted feature vector is input into the trained support vector machine model and the particle swarm optimization support vector machine model. By extracting the sample entropy of the load spectrum of the disc cutter as the feature vector, the particle swarm optimization support vector model is used to identify the cutting state of the coal and rock. The recognition accuracy of the support vector machine model before and after the improvement is compared and analyzed. The results show that compared to the unoptimized support vector machine, the support vector machine optimized by particle swarm optimization can identify the load spectrum of the coal more quickly and accurately. The recognition accuracy is 96,82%, which verifies the effectiveness of the particle swarm optimization support vector machine model in identifying the load spectrum of the coal and rock disc cutter.

Margin Maximization of Text Classification based on Support Vector Machine

Abstract: This project focuses on developing a margin maximization model for topic categorization using Support Vector Machines (SVM). Topic categorization aims to classify text documents into predefined topic categories. The proposed model leverages SVM's ability to create a hyperplane with maximum margin between different topics, enhancing the classification performance. The project begins with dataset preparation, where a labeled dataset covering a wide range of topics is gathered. The text data undergoes preprocessing, including cleaning, normalization, and conversion into numerical representations. Feature extraction techniques such as TF-IDF or word embeddings are employed to capture important features related to topics. The dataset is split into training and testing sets, with the training set used to train the SVM model. The SVM model is trained with a focus on maximizing the margin, utilizing techniques like soft-margin SVM or the kernel trick to handle non-linear separable data. Hyperparameters of the SVM model are tuned to optimize its performance. The trained SVM model is then used to predict the topics of the testing data, and its performance is evaluated using standard metrics such as accuracy, precision, recall, and F1 score. The findings of this project demonstrate the effectiveness of the margin maximization model for topic categorization using SVM

Hardware fingerprint construction of optical transmitters based on permutation entropy spectrum of chaos for secure authentication

In this paper, an optical transmitter authentication method using hardware fingerprints based on the permutation entropy spectrum (PES) of electro-optic chaos is proposed. By means of permutation analysis of chaotic time series generated by the electro-optic feedback loop, PES is defined and used as the hardware fingerprint of optical transmitters for secure authentication. The time division multiplexing (TDM) module is used to combine the chaotic signal and the message, while the optical temporal encryption (OTE) module is introduced to ensure the security of the fingerprint. At the receiver, the time division demultiplexing (TDD) module and the optical time decryption (TDD) module are introduced to recover the original chaotic signal. The trained support vector machine (SVM) models will then recognize legal and illegal optical transmitters by judging whether the PES of the received chaotic signal is legitimate or not. The simulation and experimental results show that PES of electro-optic chaos has obvious fingerprint characteristics and is sensitive to the feedback delay T in the feedback loop. The authentication system based on PES can recognize both legal and illegal transmitters in optical networks, and the recognition accuracy can reach about 97%. Under the condition of a low signal-to-noise ratio, the authentication system based on the permutation entropy spectrum can still distinguish legal and illegal transmitters effectively. Meanwhile, our scheme has high controllability and flexibility.

A Filter-SQP strategy for training Support Vector Machine models

This paper introduces a filtering strategy for addressing optimization problems arising in binary Support Vector Machine classification. The training optimization problem aims to solve the dual formulation which involves a quadratic objective function subjected to a linear and box constraints. Our approach employs a Filter algorithm with Sequential Quadratic Programming iterations that minimize the quadratic Lagrangian approximations. Notably, we utilize the exact Hessian matrix in our numerical experiments to seek the desired classification function. Moreover, we present a Filter algorithm combined with the Augmented Lagrangian method aiming to accelerate the algorithm convergence. To substantiate our method's effectiveness, we conduct numerical experiments through MATLAB, comparing outcomes with alternative methodologies detailed in existing literature. Numerical experiments shows that the Filter--SQP combined with Augmented Lagrangian method is competitive and efficient method compared with an interior-point based solver and LIBSVM software in relation of classification metrics and CPU-time.

Characterization and discrimination of the flavor profiles of Chinese indigenous sheep breeds via electronic sensory, smart instruments and chemometrics

This study was conducted to analyze the flavor profile of six commercially available Chinese indigenous breeds of sheep meat and to distinguish them based on their flavor characteristics. Various techniques such as gas chromatography-mass spectrometry (GC-MS), electronic nose (E-nose), high performance liquid chromatography (HPLC) and electronic tongue (E-tongue) were employed to analyze both volatile and non-volatile compounds in the meat samples. The results showed that, a total of 28 volatile compounds in sheep meat were detected, with 1-octen-3-ol, 1-pentanol, hexanal, octanal, and nonanal identified as the major contributors to its aroma. Tan Sheep and Oula Sheep had significantly higher concentrations of 1-pentanol, nonanal, 1-octen-3-ol, 1-octanol, and hexanoic acid than other breeds. Sweet amino acids were the most abundant free amino acids, comprising 72–85% of the total FAAs, and inosinic acid was the most abundant ATP-related compound. Small-Tailed Han sheep had significantly higher levels of SFAAs, total FAAs, and AMP, while Tan sheep had significantly higher levels of UFAAs, and Oula Sheep had significantly higher levels of GMP. The response values of differential electronic sensors W5S, W1W, NMS, PKS, and SCS were found to be useful in discriminating between breeds of sheep meat. The fusion electronic sensory signals were used to train support vector machine and random forest models, both achieving 100% accuracy in discriminating between sheep breeds.

Fingerprint construction of optical transmitters based on the characteristic of electro-optic chaos for secure authentication.

In this paper, an optical transmitter authentication method using hardware fingerprints based on the characteristic of electro-optic chaos is proposed. By means of phase space reconstruction of chaotic time series generated by an electro-optic feedback loop, the largest Lyapunov exponent spectrum (LLES) is defined and used as the hardware fingerprint for secure authentication. The time division multiplexing (TDM) module and the optical temporal encryption (OTE) module are introduced to combine chaotic signal and the message to ensure the security of the fingerprint. Support vector machine (SVM) models are trained to recognize legal and illegal optical transmitters at the receiver. Simulation results show that LLES of chaos has the fingerprint characteristic and is highly sensitive to the time delay of the electro-optic feedback loop. The trained SVM models can distinguish electro-optic chaos generated by different feedback loops with a time delay difference of only 0.03ns and have a good anti-noise ability. Experimental results show that the recognition accuracy of the authentication module based on LLES can reach 98.20% for both legal and illegal transmitters. Our strategy can improve the defense ability of optical networks against active injection attacks and has high flexibility.

An analytical identification model for ancient glass based on support vector machine and cluster analysis

Glass artifacts are easily weathered by environmental influences during burial, and the internal chemical composition is exchanged with the external environment, resulting in changes in its composition ratio. For archaeological interrogation, it is of great significance to study the chemical composition of glass before and after weathering and related issues. In this paper, a support vector machine model is used to reasonably separate the data points of high potassium glass and lead-barium glass, and a linear classification function is derived, i.e., the classification law of the two types of glass is obtained. Then, this paper performs R-type cluster analysis on the chemical composition indexes first, and reduces the dimensionality of several indexes; then performs Q-type cluster analysis on the sample detection points, and performs a one-way ANOVA test on the classification scheme of different numbers of subclasses obtained, and finally divides them into four subclasses. On this basis, the data of the unknown class of glass artifacts are substituted into the trained support vector machine model, and the type to which they belong can be determined by calculating the value of the judgment linear classification function.

Flood Monitoring in the Middle and Lower Basin of the Yangtze River Using Google Earth Engine and Machine Learning Methods

Under the background of intensified human activities and global climate warming, the frequency and intensity of flood disasters have increased, causing many casualties and economic losses every year. Given the difficulty of mountain shadow removal from large-scale watershed flood monitoring based on Sentinel-1 SAR images and the Google Earth Engine (GEE) cloud platform, this paper first adopted the Support Vector Machine (SVM) to extract the water body information during flooding. Then, a function model was proposed based on the mountain shadow samples to remove the mountain shadows from the flood maps. Finally, this paper analyzed the flood disasters in the middle and lower basin of the Yangtze River (MLB) in 2020. The main results showed that: (1) compared with the other two methods, the SVM model had the highest accuracy. The accuracy and kappa coefficients of the trained SVM model in the testing dataset were 97.77% and 0.9521, respectively. (2) The function model proposed based on the samples achieved the best effect compared with other shadow removal methods with a shadow recognition rate of 75.46%, and it alleviated the interference of mountain shadows for flood monitoring in a large basin. (3) The flood inundated area was 8526 km2, among which, cropland was severely affected (6160 km2). This study could provide effective suggestions for relevant stakeholders in policy making.

Study on wearable pregnancy diagnosis device for rabbits based on photoelectric sensor

Pregnancy diagnosis is essential for rabbits' reproductive management, which allows for re-insemination, and reduces the laboring interval in commercial operations. This study aimed to develop a wearable, non-invasive pregnancy identification device for rabbits. A sensing probe contained three LEDs (Light wavelength is 660 nm, 850 nm, and 930 nm respectively), two Si-based photodiodes, and a host with signal processing was developed. Room-temperature-vulcanized (RTV) silicone rubber, aluminum oxide powder, and dyes of different colors were used to make the abdominal tissue phantom of pregnant female rabbits (PRs) and non-pregnant female rabbits (NPRs) to test the sampling stability of the developed device. The results showed that the variation coefficient of the sampling data of the device was<0.013, and the stability met the use requirements. Sampling data from 419 female rabbits were used to build partial least squares discrimination analysis (PLS-DA) and support vector machine (SVM) models of different abdominal hair densities, including dense hair, sparse hair, hair removal, and hair mixing. The classification results of the two models showed that SVM had the best performance, the Ac, Re, Pre, and F-score are more than 80%, respectively. Then, the trained SVM model was then used as the identification model and downloaded to the developed device. Another 218 rabbits were used to verify the performance of the developed device, and the prediction performance was the best among the hair removal rabbits, with Ac, Re, Pre, and F-socre of 83.94%, 83.67%, 81.19%, and 82.41% respectively. The results show that the wearable device developed in this study has a good prediction performance for rabbits with abdominal hair removal, and provides a new idea for pregnancy diagnosis of rabbits.

Drive-by bridge damage detection using Mel-frequency cepstral coefficients and support vector machine

Bridge damage detection using vibration data has been confirmed as a promising approach. Compared to the traditional method that typically needs to install sensors or systems directly on bridges, the drive-by bridge damage detection method has gained increasing attention worldwide since it just needs one or a few sensors instrumented on the passing vehicle. Bridge frequencies extracted from the vehicle’s vibrations can be good references for damage detection. However, extant literature considered mainly low-frequency responses of the vehicle, while the high-frequency responses that also contained the bridge’s damage information were often ignored. To fill this gap, this paper developed a damage detection approach that utilized both low and high-frequency responses of the passing vehicle. Mel-frequency cepstral coefficients (MFCCs) and support vector machine (SVM) were employed to classify damage severity. Firstly, the vehicle’s frequency responses are utilized as input features to train SVM models to identify the bridge’s condition. Then, to reduce dimensions of inputs and improve training efficiency, frequency responses are projected from the Hertz scale into the Mel scale, and two means using MFCCs are used to feed different SVM models. A laboratory experiment with a U-shaped continuous beam and a model car was used to verify the effectiveness of the proposed method. Results showed that high-frequency responses contain much information about the bridge’s conditions, and using MFCCs could apparently improve computational efficiency. The errors of damage detection when a heavy car was employed were within 5%.

Evaluation of four machine learning models for signal detection.

Logistic regression-based signal detection algorithms have benefits over disproportionality analysis due to their ability to handle potential confounders and masking factors. Feature exploration and developing alternative machine learning algorithms can further strengthen signal detection. Our objective was to compare the signal detection performance of logistic regression, gradient-boosted trees, random forest and support vector machine models utilizing Food and Drug Administration adverse event reporting system data. Cross-sectional study. The quarterly data extract files from 1 October 2017 through 31 December 2020 were downloaded. Due to an imbalanced outcome, two training sets were used: one stratified on the outcome variable and another using Synthetic Minority Oversampling Technique (SMOTE). A crude model and a model with tuned hyperparameters were developed for each algorithm. Model performance was compared against a reference set using accuracy, precision, F1 score, recall, the receiver operating characteristic area under the curve (ROCAUC), and the precision-recall curve area under the curve (PRCAUC). Models trained on the balanced training set had higher accuracy, F1 score and recall compared to models trained on the SMOTE training set. When using the balanced training set, logistic regression, gradient-boosted trees, random forest and support vector machine models obtained similar performance evaluation metrics. The gradient-boosted trees hyperparameter tuned model had the highest ROCAUC (0.646) and the random forest crude model had the highest PRCAUC (0.839) when using the balanced training set. All models trained on the balanced training set performed similarly. Logistic regression models had higher accuracy, precision and recall. Logistic regression, random forest and gradient-boosted trees hyperparameter tuned models had a PRCAUC ⩾ 0.8. All models had an ROCAUC ⩾ 0.5. Including both disproportionality analysis results and additional case report information in models resulted in higher performance evaluation metrics than disproportionality analysis alone.

Evolutionary computation-based machine learning for Smart City high-dimensional Big Data Analytics

Science and technology development promotes Smart City Construction (SCC) as a most imminent problem. This work aims to improve the comprehensive performance of the Smart City-oriented high-dimensional Big Data Management (BDM) platform and promote the far-reaching development of SCC. It comprehensively optimizes the calculation process of the BDM platform through Machine Learning (ML), reduces the dimension of the data, and improves the calculation effect. To this end, this work first introduces the concept of SCC and the BDM platform application design. Then, it discusses the design concept of using ML technology to optimize the calculation effect of the BDM platform. Finally, the Tensor Train Support Vector Machine (TT-SVM) model is designed based on dimension reduction data processing. The proposed model can comprehensively optimize the BDM platform, and the model is compared with other models and evaluated. The research results show that the accuracy of the reduced dimension classification of the TT-SVM model is more than 95. The lowest average processing time for the model’s reduced dimension classification is about 1ms. The model’s highest data processing accuracy is about 98%, and the average processing time is between 1.0–1.5ms. Compared with traditional models and BDM platforms, the proposed model has a breakthrough performance improvement, so it plays an important role in future SCC. This work has achieved a great breakthrough in big data processing, and innovatively improved the application mode of high-dimensional big data technology by integrating multiple technologies. Therefore, the finding provides targeted technical reference for algorithms in BDM platform and contributes to the construction and improvement of Smart City.

Adaptive SVM-based pixel accumulation technique for a SPAD-based lidar system.

This paper proposes an adaptive support vector machine (SVM)-based pixel accumulation technique for the single-photon avalanche diode (SPAD)-based flash lidar system to greatly improve the accuracy of depth images. In this method, an adaptive incremental SVM classifier is proposed to distinguish the target and background to avoid the boundary blur caused by accumulation. With the SVM classifier, only the pixels belonging to the same target are accumulated. Then by accumulating the multiple measurements of the results of neighboring pixels, the influence of noisy photons can be reduced, and an accurate peak position of the laser signal in the histogram generated by the SPAD detector can be derived. Different from most existing classification algorithms in which the SVM model is used to recognize the target in depth, the image is trained with huge amounts of offline data, which neglect the ability to handle the changes of target's features by building a small subset of input data that contains all needed information of the trained SVM model. The new method could quickly update itself with the motion of target or light intensity changes online and definitely enhance the practicability of the proposed depth image-processing method. Experimental results show that using the proposed adaptive pixel accumulation technique, the mean squared error of depth images can be reduced by 61% compared with raw images and 44% compared with the static SVM-based method when the target moves.

Identifying the 3FHL Catalog. VI. Swift Observations of 3FHL Unassociated Objects with Source Classification via Machine Learning

The Third Catalog of Hard Fermi Large Area Telescope Sources (3FHL) reports the detection of 1556 objects at E > 10 GeV. However, 177 sources remain unassociated and 23 are associated with a ROSAT X-ray detection of unknown origin. Pointed X-ray observations were conducted on 30 of these unassociated and unknown sources with Swift−XRT. A bright X-ray source counterpart was detected in 21 out of 30 fields. In five of these 21 fields, we detected more than one X-ray counterpart, totaling 26 X-ray sources analyzed. Multiwavelength data was compiled for each X-ray source detected. We find that 21 out of the 26 X-ray sources detected display the multiwavelength properties of blazars, while one X-ray source displays the characteristics of a Galactic source. Using trained decision tree, random forest, and support vector machine models, we predict all 21 blazar counterpart candidates to be BL Lacertae objects (BL Lacs). This is in agreement with BL Lacs being the most populous source class in the 3FHL.

A machine learning‐based approach for dielectric strength prediction of long air gaps with engineering configurations

It is a long-term goal in external insulation studies to determine the discharge voltages of complicated engineering gaps by simulation methods. Based on the one-to-one correspondence between air gap structure and the static electric field (EF) distribution, this paper characterizes the transmission tower gap configuration by spatial EF features, which were used for machine learning to achieve switching impulse discharge voltage prediction. An interelectrode path and a conical zone between the energized sub-conductor and the crossarm or the tower window were considered as EF regions strongly associated with gap breakdown, where 73 parameters were extracted to construct a feature set. Taking 15 extra-high voltage (EHV) transmission tower gaps as training samples, with different gap distances and tower configurations, their EF features were input to a support vector machine (SVM) for model training to establish the relationships with discharge voltages. The trained SVM model was used to predict the impulse discharge voltages of 20 EHV and ultra-high voltage (UHV) transmission tower gaps. The prediction results with different feature dimensions and various sizes of conical zones were compared to the experimental values, which demonstrate similar variation trends and acceptable errors. This study contributes to realize insulation strength calculation of engineering gaps.