Average Relative Prediction Error Research Articles

Elemental quantitation and evaluation of hydrocarbon in shale using fiber-optic laser induced breakdown spectroscopy

To meet the demands of in situ analysis, a fiber-optic laser-induced breakdown spectroscopy (FO-LIBS) system was established to predict the elemental concentrations and pyrolytic parameters. The key parameters, including pulse energy, gate delay, type of surrounding gas, and flow rate of gas, were optimized to improve the signal-to-noise ratio of the spectral lines, and the evolution characteristics of the plasma morphology in different gas atmospheres were investigated. Under the optimized experimental conditions, C, H, Si, Ca, Fe, and Mg were calibrated using pellet samples based on partial least squares regression, and the average relative prediction error of the natural samples was lower than 10%. In addition, the support vector regression method was utilized to predict pyrolytic parameters such as TOC, Pg, and Tmax with the R2 of all calibration curves being higher than 0.97. The abundance and generation potential of hydrocarbons were evaluated using the predicted TOC and Pg, and Tmax could contribute to the thermal maturity of the kerogen.

Texture feature similarity-based roughness intelligent evaluation: a case study applied to milled surfaces

Surface roughness is of great significance in maintaining mechanical performance and improving the reliability of the equipment. However, fast surface roughness evaluations that are sufficiently stable and efficient for engineering in situ use have not yet been realized. To address this issue, an image-driven roughness intelligent method is proposed in this research. By evaluating the texture similarity intelligently between the testing image and the reference image, the surface roughness of the testing image can be acquired. Firstly, with a proposed adaptive texture extraction method, the texture feature of an image can be extracted even under a complex background. Secondly, by establishing the graph structure of the texture grayscale features, the similarity between different images is evaluated. Finally, by establishing a sparrow-optimized support vector machine regression method, the mapping relationship between the similarity and the surface roughness can be acquired. The experimental results indicate that the proposed method for intelligent evaluation of roughness has superior prediction performance (the average relative prediction error of Ra and Rz are 8.8156% and 8.0571%, respectively). Therefore, this work provides a useful tool for non-contact detection of workpiece surface roughness.

Development of a yield monitoring model based on analysis of surveys and images of the field

The study is devoted to the construction of a mathematical model of the yield of agricultural crops, which includes three components: trend, seasonal, and random. The developed model shows the dependence of yield on phenological indicators, the quality of land resources, management efficiency, and other random factors. The trend and seasonal components of the yield model do not depend on random factors and can therefore be used to predict yield. It is proposed to consider the trend component of the model as a piecewise linear function, and the seasonal component of the model as a linear harmonic regression. A method for analyzing multispectral images with consideration of geoinformation data has been developed to assess phenological indicators. This method includes determining the threshold value using Otsu method to find the density of the agricultural crop in the field. Data on crop density, supplemented with geodata about the plot boundaries, are used to calculate the yield. A comparison of yield forecasts for three crops in the Chernihiv region was made using observations of the phenological indicators of crops throughout the year and over 3 months. It was found that yield is significantly determined by plant development in the first months after germination. The comparison of yield forecasts was made with data from the State Statistics Service of Ukraine and forecasts made using the WOFOST simulation model. It was established that the average relative error of yield prediction using the developed model is 2.96% when using observations of the phenological indicators of crops throughout the year and 4.51% when observing over 3 months. This accuracy is sufficient and comparable to the average accuracy of predictions based on the WOFOST model, which is 3.62%).

Determination of phenolic compounds in water using a multivariate statistical analysis method combined with three-dimensional fluorescence spectroscopy.

Phenolic compounds are toxic chemical pollutants present in water. Three-dimensional fluorescence spectroscopy analysis is an effective and rapid method for real-time phenol monitoring in aquatic environments. However, similar chemical structures of phenols result in highly overlapping three-dimensional fluorescence spectra. Therefore, it is extremely difficult to analyze and quantify the concentration of components in a mixture system that includes two or more phenolic compounds. In this article, we study the mixed phenol system containing phenol, o-cresol, p-cresol, m-cresol, catechol, and resorcinol combined with excitation-emission matrix (EEM) fluorescence data. A multivariate statistical method called best linear unbiased prediction (BLUP) is proposed to analyze the spectra with the aim to achieve quantitative results and a trilinear decomposition algorithm called parallel factor analysis (PARAFAC) was used for comparison. Two experiments with different calibration samples were set to validate the effectiveness of BLUP through recovery, ARecovery (Average Recovery), AREP (Average Relative Error of Prediction), and RMSE (Root Mean Square Error). Overall, the average recovery of each component in experiment 1 and experiment 2 ranged from 95.91% to 111.62% and 82.91% to 129.02%, respectively. Based on the results of the experiments, the concentration of phenolic compounds in water can be quantitatively determined by combining three-dimensional fluorescence spectroscopy with the BLUP method.

AS-NeSt: A Novel 3D Deep Learning Model for Radiation Therapy Dose Distribution Prediction in Esophageal Cancer Treatment With Multiple Prescriptions

Purpose: Implementing artificial intelligence (AI) technologies allows for the accurate prediction of radiotherapy dose distributions, enhancing treatment planning efficiency. However, esophageal cancers present unique challenges due to tumor complexity and diverse prescription types. Additionally, limited data availability hampers the effectiveness of existing AI models. This study developed a deep learning model, trained on a diverse dataset of esophageal cancer prescriptions, to improve dose prediction accuracy. Methods and Materials: We retrospectively collected data from 530 esophageal cancer patients, including single-target and simultaneous integrated boost (SIB) prescriptions, for model building. The proposed Asymmetric ResNeSt (AS-NeSt) model features novel 3D ResNeSt blocks and an asymmetric architecture. We constructed a loss function targeting global and local doses and validated the model's performance against existing alternatives. Model assisted experiments were used to validate its clinical benefits. Results: The AS-NeSt model maintained an absolute prediction error below 5% for each dosimetric metric. The average dice similarity coefficient (DSC) for isodose volumes was 0.93. The model achieved an average relative prediction error of 2.02%, statistically lower than HD-Unet (4.17%), DoseNet (2.35%), and DCNN (3.65%). It also demonstrated significantly fewer parameters and shorter prediction times. Clinically, the AS-NeSt model raised physicians’ ability to accurately pre-assess appropriate treatment methods before planning from 95.24% to 100%, reduced planning time by over 61% for junior dosimetrists and 52% for senior dosimetrists, and decreased both inter- and intra-dosimetrist discrepancies by more than 50%. Conclusion: The AS-NeSt model, developed with innovative 3D ResNeSt blocks and an asymmetric encoder-decoder structure, has been validated using clinical esophageal cancer patient data. It accurately predicts 3D dose distributions for various prescriptions, including SIB, showing potential to improve the management of esophageal cancer treatment in a clinical setting.

Dynamic weighing algorithm for dairy cows based on time domain features and error compensation

Weight is an essential indicator for the growth and development of cows. Traditional static weighing method requires a lot of labor and financial resources, the existing dynamic weighing algorithm has no consideration of the influence of hidden information on weighing results during cows' movement, and fails to meet the weighing requirements in complex environments. In this paper, we propose a cow dynamic weighing algorithm based on time domain features and error compensation, quantitatively analyze the hidden information in the dynamic weighing signal, and compensate for the error of the preliminary prediction results, leading to a high-precision dynamic weighing of cows. Firstly, the dynamic weighing signal of cows is pre-processed to obtain the effective signal, and each time domain feature of the effective signal are extracted. Then, the effective signal is decomposed by using the empirical wavelet transform (EWT), and the trend component's mean value is taken as the prediction value of EWT, so that the error between the prediction weight value and the real weight can be calculated. Next, four time-domain features are used as input and the error values are used as output to construct the data set, and the sparrow search algorithm (SSA) is used to obtain the optimal parameters of support vector regression (SVR) model for the training. Finally, the error prediction values output from the model are summed with the prediction values of EWT as the final weight prediction value. To evaluate the performance of the method in this paper, we collect dynamic weighing data from 626 cows, and then randomly select 501 data as the training set while the remaining 125 data as the test set. The average relative error of weight prediction of our method is 0.68%, and the root mean square error is 4.83 kg, better than the existing dynamic weighing algorithms. The experimental results show that the method in this paper is feasible to quantitatively analyze the implicit information of the effective signal according to the time domain features, and the error compensation method can effectively reduce dynamic weighing errors, particularly under extreme conditions, making the method more robust and better meeting the practical needs of farms.

Sustainable development of vocal music ecology in a digital ecological environment

The phenomenon of noise interference during vocal transmission can lead to the problem of poor vocal transmission quality. This paper proposes a study on the sustainability of vocal ecology in a digital ecological environment. First, the matching tracking algorithm can extract the time-frequency characteristics of the effective signal, attenuate the interference of noise, and improve the propagation quality. A sustainable development GA-BP network model is established, and the adjustment amount of each weighting coefficient is obtained according to the gradient algorithm and using the inertia adjustment strategy. The coordination is regulated through feedback control strategies to ultimately achieve ecological sustainability of vocal music. The analysis results show that the average relative error of the simulation prediction of the sustainable development GA-BP network model is 3.54%, the maximum relative error is 8.11%, and the average relative error is within 5% of 70%. It has significant superiority and high efficiency in comparison with the prediction degree of the traditional model.

DEM simulation of impact charge on polymer particles by single and repeated impact on different metal plates

Charge transfer between the particle and wall occurs after each collision in powder handling operations, called "contact electrification". In the present study, a LIGGGHTS code was developed by adding the Hertzian contact area formula and condenser model to predict the particle charge after every impact. The simulation results showed that, the particle with a higher impact velocity became saturated under fewer collisions. Considering that the charging constant depends on the impact velocity and initial particle charge, the Average Relative Error (ARE) of simulation prediction significantly decreased from 22 and 16% to 10.6 and 11% for Rubber particles at impact velocities of 4.8 and 2.8 m/s, respectively. Therefore, the current modified LIGGGHTS code has a great potential to predict the impact charge of particle-wall collision.

Selectivity Quantification with the Fluorescent Quantitative Model-Assisted Semi-Selective Probe (Carbon Dots): Accurate Determination of Chlortetracycline in Aqueous Environments with Interference.

Probes such as carbon dots (C-dots) have extensive and important applications in the quantitative analysis of complex biological and environmental systems. However, the development of probes is often hindered by incomplete selectivity, i.e., a probe that responds to one substance is also prone to respond to coexisting structurally similar substances. Therefore, the above dilemma often leads to be developed as semi-selective probes, so that the development of probes is abandoned halfway. This work shows how a semi-selective probe can enhance selectivity by combining a proper multivariate calibration model. Primarily, we developed a semi-selective fluorescent probe that responded to tetracyclines (TCs) with discarded tobacco leaves. Then, we introduced the multivariate quantitative fluorescence model (QFM) to enhance its selectivity and solve the problem of fluorescence spectral shift. For the determination of chlortetracycline (CTC) with this semi-selective C-dots probe in mineral and lake water samples and compared to the traditional quantitative model, the introduced QFM resulted in an average relative predictive error (ARPE) in mineral water spiked samples decreased from 57.1 to 5.6%, which reduced the ARPE in the lake water spiked samples from 18.1 to 4.7%. The above results show that the QFM-assisted semi-selective probe C-dots strategy (QFMC-dots) can enhance selectivity, and QFMC-dots achieved high-selective and accurate determination of CTC in interfering mineral and lake water samples, with the limit of detection and limit of quantitation of 0.55 and 1.66 μM, respectively. The proposed strategy of enhancing selectivity by introducing a proper multivariate calibration model can reduce the difficulty and increase success rate of developing probes, which can be expected to provide an interesting alternative for the development of probes, especially when encountering semi-selective problems.

Accurate Cutting-Force Measurement with Smart Tool Holder in Lathe.

Cutting force in lathe work is closely related to tool wear and affects the turning quality. Direct measurement of the cutting force by measuring the strain of the tool holder is challenging because the tool holder design aims to be highly rigid in order to undertake large cutting forces. Accordingly, the most popular dynamometer designs modify the standard tool holder by decreasing the structural rigidity of the holder, which reduces the machining precision and is not widely accepted. In order to solve the issue of the low stiffness of the dynamometer reducing the machining precision, in this paper, the ultra-low strain on the tool holder was successfully detected by the highly sensitive semiconductor strain gauges (SCSG) adjacent to the blade cutting insert. However, the cutting process would generate much heat, which increases the force measuring area temperature of the tool holder by about 30 °C. As a result, the readout drifted significantly with the temperature changes due to the high temperature coefficient of SCSG. To solve this problem, the temperature on the tool holder was monitored and a BP neural network was proposed to compensate for temperature drift errors. Our methods improved the sensitivity (1.14 × 10-2 mV/N) and the average relative error of the BP neural network prediction (≤1.48%) while maintaining the original stiffness of the tool holder. The smart tool holder developed possesses high natural frequency (≥6 kHz), it is very suitable for dynamic cutting-force measurement. The cutting experiment data in the lathe work show comparable performance with the traditional dynamometers and the resolution of the smart tool holder is 2 N (0.25% of total range).

Tensile properties and prediction of DP780 treated with rare earth Ce based on DE-SVM-Ⅱ

In order to solve the weak tensile properties of DP780 welded joints, rare earth Ce was added to improve the tensile properties of DP780 welded joints. In this paper, the microstructure of DP780 was characterized by metallographic microscopy, scanning electron microscopy and energy spectrum analyzer. An improved differential evolution algorithm optimized support vector machine prediction model (DE-SVM-II) was established to predict the tensile properties of DP780. The results show that when the content of Ce was 0.05%, the tensile strength increased by 378 MPa, the microstructure of molten metal was uniform and the grain boundary of austenite was tight. In the heat affected zone, the size of ferrite decreases, the fracture morphology shows equiaxial depression, and the fiber zone accounts for a large proportion. In addition, the fitness of the mean square error (MSE) of the tensile strength curve in the DE-SVM-Ⅱ prediction was lower than that of the SVM prediction error, and begin to converge in the 10th generation. The average relative error of the DE-SVM model prediction of elongation was within 3%. Compared with the predicted value of BP neural network, the error of predicting tensile strength and elongation of DE-SVM-II was reduced by 1.36% and 15.10%, respectively, which had high stability and accuracy.

Strength Prediction of Ball-Milling-Modified Phosphorus Building Gypsum Based on NSGM (1,4) Model.

Phosphogypsum is an industrial byproduct from the wet preparation of phosphoric acid. Phosphorus building gypsum can be obtained from phosphogypsum after high-thermal dehydration. This study aimed to analyze the influence of ball milling with different parameters on the strength of phosphorus building gypsum. In this paper, the absolute dry flexural strength and the absolute dry compressive strength of phosphorus building gypsum were compared under different mass ratios of material to ball, ball-milling speed, and ball-milling time, and the NSGM (1,4) model was applied to model and predict the strength of phosphorus building gypsum modified by ball milling. According to the research results, under the same mass ratio of material to ball and ball-milling speed, the absolute dry flexural strength and absolute dry compressive strength of phosphorus building gypsum firstly increased and then decreased with the increase in milling time. The NSGM (1,4) model established in this paper could effectively simulate and predict the absolute dry flexural strength and the absolute dry compressive strength of the ball-milling-modified phosphorus building gypsum; the average relative simulation errors were 12.38% and 13.77%, and the average relative prediction errors were 6.30% and 12.47%.

Accurate varietal classification and quantification of key quality compounds of grape extracts using the absorbance-transmittance fluorescence excitation emission matrix (A-TEEM) method and machine learning

Rapid and accurate quantification of grape berry phenolics, anthocyanins and tannins and identification of grape varieties are both important for effective quality control of harvesting and initial processing for winemaking. Current reference technologies, including High-Performance Liquid Chromatography (HPLC), can be rate-limiting and too complex and expensive for effective field operations. In this paper, we analyse robotically prepared grape extracts from several key varieties (n = Calibration/n = Prediction samples), including Cabernet-Sauvignon (64/10), Grenache (16/4), Malbec (14/4), Merlot (56/10), Petite Sirah (52/10), Pinot noir (54/8), Syrah (20/2), Teroldego (14/2) and Zinfandel (62/12). Key phenolic and anthocyanin parameters measured by HPLC included Catechin, Epicatechin, Quercetin Glycosides, Malvidin 3-glucoside, Total Anthocyanins and Polymeric Tannins. Split samples diluted 50-fold in 50 % EtOH pH 2 were analysed in parallel using the A-TEEM method following Multi-block Data Fusion of the absorbance and unfolded EEM data. A-TEEM chemical data were calibrated (n = 390) using Extreme Gradient Boosting (XGB) Regression and evaluated based on the Root Mean Square Error of the Prediction (RMSEP), the Relative Error of Prediction (REP) and Coefficient of Variation (R2P) of the Prediction data (n = 62). The regression results yielded an average Relative Error of Prediction (REP) of 5.89 ± 2.47 % and an R2P of 0.941 ± 0.025. While we consider the REP values to be in the acceptable range at significantly < 10 %, we acknowledge that both the grape extraction method repeatability and HPLC reference method sample repeatability (5-8 % RSD) likely constituted the major sources of variation compared to the A-TEEM instrumental sample repeatability (< 2 % RSD). The varietal classification was analysed using Agglomerative Hierarchical Cluster Analysis (HCA) and XGB discrimination analysis of the multi-block data. The classification results yielded 100 % True Positive and True Negative responses for the Calibration and Prediction Data for all tested varieties. We conclude that the A-TEEM method requires a minimum of sample preparation and rapid acquisition times (< 1 min) and can serve as an accurate secondary method for both grape varietal identification and phenolic quantification. Importantly, the software application of the regression and classification models can be effectively automated for operators.

An Integration Model for Flux Density Distribution Formed by a Heliostat

An accurate flux density calculation is essential for optimizing and designing solar tower systems. Most of the existing methods introduce multiple assumptions, and the accuracy and scope of the application are limited. This paper proposes an integration model used to calculate the flux density distribution after only applying the Gaussian model for solar brightness distribution. It is the first time that multiple reflections and the influence of the optical error transferred from different planes of the glass mirror are considered in order to build an optical model for the flux density of a heliostat. The reflection from two surfaces of the glass mirror used to form three main parts of beams was considered in the present model, and Fresnel’s equations were applied to calculate the energy of the three parts of reflected rays. An elliptic Gaussian model was applied for the optical error distribution of the heliostat. The model error was evaluated using the experimental data of ten heliostats, and the applicability and accuracy of the model were verified through flux distribution and an intercept factor. The average relative prediction error of the present model from the experimental data was only 2.83%, which is less than SolTrace and other models.

Research on Nonlinear Time Series Processing Method for Automatic Building Construction Management

Aiming at the nonlinear time series of automatic building construction management, a neural network prediction model is proposed to analyze and process the nonlinear sequence of deformation monitoring number cutter. The specific content of this method is as follows: for the noise problems existing in deformation monitoring data, a wavelet is used to denoise the preprocessing; for the BP network and RBF network commonly used in neural networks, the performance of the two networks is compared and demonstrated by MATLAB program, which proves that RBF neural network can significantly improve the accuracy of deformation prediction. By comparing the results, the maximum relative error of BP network prediction is 18.59%, while the maximum relative error of RBF network prediction is 29.16%, and the average relative error of 13P network prediction is 7.02%, while the average relative error of RBF network prediction value is 10.5%. The comprehensive error of network prediction is 6.1%, RBF network prediction is 8.52%, the standard deviation RMSE of BP network prediction error is 15.347, and that of RBF network prediction error is 21.401, and it shows that the prediction accuracy of BP network is higher than that of RBF network.

Proper orthogonal decomposition of continuum-dominated emission spectra for simultaneous multi-property measurements

A novel spectrum calibration method is proposed for high-accuracy emission spectroscopy to extract quantitative gas property information from the spontaneous photon emission spectra. This is challenging, particularly in the presence of broadband emission, which smears the characteristic local spectral features that are often used as property indicators in conventional emission spectroscopy. Proper orthogonal decomposition (POD) can reduce the dimension (ND) of the broadband spectrum profile by selecting a number (NM≪ND) of dominant bases that span the gas-property-dependent spectrum space. Each POD-decomposed spectral basis represents critical broadband spectral features that are sensitive to the gas properties of interest. A reduced-order model (ROM) employing the kriging method correlates the entire emission spectrum profile that decomposed by the bases to multiple gas properties. To provide a spectrum database to train the ROM, chemiluminescence was captured from a flat flame with varied diluent gas compositions (N2 and CO2), fuel-oxidizer equivalence ratios, and pressures (1–5 bar). Subsequently, the ROM employing an emission spectrum can predict the gas properties under high-temperature and high-pressure conditions where conventional sensors are inapplicable. The average relative prediction errors obtained using this new analysis method were approximately 3.2% and 1.3% for the pressure and equivalence ratio, respectively.

Solving 2D Poisson's equation based on conditional generative adversarial network

AbstractThe numerical simulation of electromagnetic problems is usually complex and time‐consuming, so a fast and accurate solver is urgently needed. In this paper, a deep learning method based on conditional generative adversarial network (CGAN) is used to solve the 2D Poisson's equation in electrostatic field. The relative permittivity distribution and source charge distribution in the two dimensional region are regarded as label of CGAN, and the potential distribution is the real image or generated image corresponding to the label. This deep learning network can be trained by reasonably designing the loss functions of generator and discriminator, and based on a large amount of samples obtained by the finite difference method. The average relative prediction error of the trained CGAN for all samples is less than 0.5%, with a significant reduction in prediction time, while that of convolution neural network (CNN) is less than 1.5%. Furthermore, the average relative prediction error of CGAN is less than 1% for the scenes different from those of training and testing samples, including triangle, square, and pentagon. The prediction results indicate that CGAN has stronger learning ability and fast prediction speed, and can be used as a fast solver for electromagnetic numerical problems.

Quantitative analysis and time-resolved characterization of simulated tokamak exhaust gas by laser-induced breakdown spectroscopy

Tokamak exhaust is an important part of the deuterium-tritium fuel cycle system in fusion reactions. In this work, we present a laser-induced breakdown spectroscopy (LIBS)-based method to monitor the gas compositions from the exhaust system in the tokamak device. Helium (He), a main impurity in the exhaust gas, was mixed with hydrogen (H2) in different ratios through a self-designed gas distribution system, and sealed into a measurement chamber as a standard specimen. A 532 nm wavelength laser pulse with an output power of 100 mJ was used for plasma excitation. The time-resolved LIBS is used to study the time evolution characteristics of the signal strength, signal-to-background ratio (SBR), signal-to-noise ratio (SNR) and relative standard deviation (RSD) of the helium and hydrogen characteristic lines. The Boltzmann two-line method was employed to estimate the plasma temperature of laser-induced plasma (LIP). The Stark-broadened profile of He I 587.56 nm was exploited to measure the electron density. From these studies, an appropriate time was determined in which the low RSD% was consistent with the high signal-to-noise ratio. The He I 587.56 nm and Hα emission lines with good signal-to-noise ratio were extracted from the spectrum and used in the external standard method and internal standard method for quantitative analysis. The test results for mixed gas showed that the average relative error of prediction was less than 11.15%, demonstrating the great potential of LIBS in detecting impurities in plasma exhaust gas.

Time-resolved spectral-image laser-induced breakdown spectroscopy for precise qualitative and quantitative analysis of milk powder quality by fully excavating the matrix information

A novel and effective method named time-resolved spectral-image laser-induced breakdown spectroscopy (TRSI-LIBS) was proposed to achieve precise qualitative and quantitative analysis of milk powder quality. To verify the feasibility of TRSI-LIBS, qualitative and quantitative analysis of milk powder quality was carried out. For qualitative analysis of foreign protein adulteration, the accuracy of models based on TRSI-LIBS was higher than those based on LIBS, with an accuracy improvement of about 5% to 10%. For the quantitative analysis of foreign protein adulteration and element content, the quantitative analysis models based on TSRI-LIBS also had better effect. For instance, limit of detection (LOD),determination coefficient of prediction (R2p), root-mean-square error of prediction (RMSEP) and average relative error of prediction (AREP) of quantitative model of calcium (Ca) content based on TRSI-LIBS improved from 1.47 mg/g, 0.95, 0.35 mg/g and 23.29% to 0.81 mg/g, 0.98, 0.20 mg/g and 12.60%.

Application of unlabelled big data and deep semi-supervised learning to significantly improve the logging interpretation accuracy for deep-sea gas hydrate-bearing sediment reservoirs

Due to the extremely complex reservoirs and strong heterogeneity, deep-sea gas hydrate logging porosity calculations still have problems, which further leads to insufficient resource calculation accuracy. Logging reservoir evaluation methods based on intelligence may be able to provide more reliable prediction results, especially the logging evaluation model based on deep learning with great potential. This paper proposes a new method to form unlabelled logging big data, and based on this, establishes a semi-supervised deep learning method suitable for deep-sea gas hydrate-bearing sediments porosity calculation, forming a porosity evaluation model. The method of forming logging big data expands 380 original data samples into 2280 labelled samples and 60050 unlabelled samples, which reduces the sampling requirements for deep-sea sediment formations with high sampling costs. The evaluation results show that the model not only obtains better results than other methods in the inspection wells corresponding to the areas where the training wells are located, but also obtains very good results in other wells that are not involved in the modelling. Compared with traditional prediction methods, the average relative error of porosity prediction is less than 4%. As far as we know, this is the first time that deep learning has been successfully applied to deep-sea hydrate sediment reservoir logging evaluation. It provides a new idea for intelligent logging evaluation of deep-sea hydrate sediment reservoirs.