Nonlinear Mapping Relationship Research Articles

Fault diagnosis method via one vs rest evidence classifier considering imprecise feature samples

The key task of fault diagnosis is to establish the nonlinear mapping relationship between fault feature sequences (FFSs) and fault modes (FMs). Therefore, it is usually necessary to detect transit jump points of FFS to separate feature samples with the different trends. These samples and the their fault labels are combined to form training samples for a fault diagnosis model (FDM). However, due to changes in fault status and uncertainty in measurement, some feature samples will imprecisely point to multiple FMs (fault labels), hardly adopted in FDM. In order to solve such imprecision problem, this paper presents an information fusion method via One vs Rest (OvR) evidence classifiers. For each FFS, the multiple OvR evidence classifiers are designed to model the imprecise relationship between the samples and FMs. Then, a two-level fusion framework is proposed to integrate the evidence of all FFSs. By using the intersection operation between One (one FM) and Rest (the other FMs), the imprecision can be significantly reduced. A rotating machinery fault diagnosis experiment is given to illustrate that the proposed method have better fault classification accuracy compared with the traditional FDMs including SVM, BPNN, KNN, FCNN, RF and Bayes.

Resolution enhancement of SOHO/MDI Magnetograms

Abstract Research on the solar magnetic field and its effects on solar power generation and space weather events has benefited from the continual improvement in instrument resolution and measurement frequency. At the same time, expanding and integrating historical observational data along the timeline is also of significant importance for studying the variations in the solar magnetic field. In astronomical data processing, super-resolution reconstruction refers to the process of using a large amount of training data to learn the nonlinear mapping relationship between low-resolution and high-resolution images in order to obtain higher-resolution astronomical images. This paper belongs to the application research of high-dimensional nonlinear regression. We use deep learning models to perform SR modeling on the SOHO/MDI magnetogram and SDO/HMI magnetogram, achieving reliable resolution enhancement of the full-disk MDI magnetogram and improving the resolution of the image to obtain more detailed information. We use a total of 9717 pairs of data that meet the selection criteria from April 2010 to February 2011 as the training set, 1332 pairs of data from March 2011 as the validation set, and 1034 pairs of data from April 2011 as the testing set. After data preprocessing, LR images of size 128*128 are randomly cropped from the MDI full-disk magnetogram, and corresponding HR images of size 512*512 are cropped from the HMI full-disk magnetogram for model training. Through testing, this study achieved reliable 4x super-resolution reconstruction of the full-disk MDI magnetogram. From the correlation coefficient, the MESR model results (0.94) are highly correlated with the HMI target magnetogram. At the same time, this method achieved the best PSNR, SSIM, MAE and RMSE scores, indicating that the MESR model can effectively reconstruct the magnetogram.

Shear wave velocity prediction based on 1DCNN-BiLSTM network with attention mechanism

The Shear wave (S-wave) velocity is an essential parameter in reservoir characterization and evaluation, fluid identification, and prestack inversion. However, the cost of obtaining S-wave velocities directly from dipole acoustic logging is relatively high. At the same time, conventional data-driven S-wave velocity prediction methods exhibit several limitations, such as poor accuracy and generalization of empirical formulas, inadequate exploration of logging curve patterns of traditional fully connected neural networks, and gradient explosion and gradient vanishing problems of recurrent neural networks (RNNs). In this study, we present a reliable and low-cost deep learning (DL) approach for S-wave velocity prediction from real logging data to facilitate the solution of these problems. We designed a new network sensitive to depth sequence logging data using conventional neural networks. The new network is composed of one-dimensional (1D) convolutional, bidirectional long short-term memory (BiLSTM), attention, and fully connected layers. First, the network extracts the local features of the logging curves using a 1D convolutional layer, and then extracts the long-term sequence features of the logging curves using the BiLSTM layer, while adding an attention layer behind the BiLSTM network to further highlight the features that are more significant for S-wave velocity prediction and minimize the influence of other features to improve the accuracy of S-wave velocity prediction. Afterward, the nonlinear mapping relationship between logging data and S-wave velocity is established using several fully connected layers. We applied the new network to real field data and compared its performance with three traditional methods, including a long short-term memory (LSTM) network, a back-propagation neural network (BPNN), and an empirical formula. The performance of the four methods was quantified in terms of their coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). The new network exhibited better performance and generalization ability, with R2 greater than 0.95 (0.9546, 0.9752, and 0.9680, respectively), RMSE less than 57 m/s (56.29, 23.18, and 30.17 m/s, respectively), and MAE less than 35 m/s (34.68, 16.49, and 21.47 m/s, respectively) for the three wells. The test results demonstrate the efficacy of the proposed approach, which has the potential to be widely applied in real areas where S-wave velocity logging data are not available. Furthermore, the findings of this study can help for a better understanding of the superiority of deep learning schemes and attention mechanisms for logging parameter prediction.

Modeling and signal integrity analysis of silicon interposer channels based on MTL and KBNN

In this paper, the equivalent circuit model of interconnect channels in silicon interposer is developed for three-dimensional integrated circuit (3-D IC) based on the theory of multi-conductor transmission line (MTL). The deep neural network (DNN) is employed to learn the nonlinear mapping relationship between the geometric parameters and the distributed parameters, and it is demonstrated that the DNN has a high prediction accuracy. In the implementation, a coarse model is firstly developed by solving the scattering parameters of the model using the chain parameter matrix. The model is optimized using a knowledge-based neural network (KBNN), and a fine model is then established. The numerical examples demonstrate that the optimized model can achieve high accuracy. Finally, the inner contour of the worst-case eye diagrams of the channel is generated using the peak distortion algorithm (PDA), with the worst eye width and height obtained.

Integrated passive design method optimized for carbon emissions, economics, and thermal comfort of zero-carbon buildings

Improving passive design methods are important for achieving zero-carbon building (ZCB) targets. Data-driven design approaches are more accurate than conventional methods for multiple-objective optimization (MOO) targets such as carbon emissions (CE), economics, and thermal comfort. In this study, a comprehensive data-driven-based PSO-SVM-NSGA-III method that assists in optimizing passive design parameters are proposed. First, the simulation results for the objectives of CE, incremental cost (IC), and time not comfort (TNC) are performed for an office building in a cold region in China using EnergyPlus and jEPlus software. Then, key influencing factors are chosen via a sensitivity analysis. Second, nonlinear mapping relationships between the passive design parameters and objectives are established with the PSO-SVM model. Compared to the BPNN, SVM, and PSO-BPNN algorithms, the PSO-SVM model exhibits superior prediction performance, with R2 values of 0.977, 0.925, and 0.903 for CE, IC, and TNC, respectively. Third, the nonlinear mapping relationships are used in the established objective functions of NSGA-II and NSGA-III. NSGA-III performed 100 iterations within 21 min and exhibited robust diversity within its population, with a hypervolume value of 0.738. Finally, the Pareto-optimal solution set is formed with the PSO-SVM-NSGA-III framework. This data-driven method offers an efficient approach for improving building performance.

Novel high-safety aeroengine performance predictive control method based on adaptive tracking weight

Increasing attention has been attracted to the dynamic performance and safety of advanced performance predictive control systems of the next-generation aeroengine. The latest research demonstrates that Subspace-based Improved Model Predictive Control (SIMPC) can overcome the difficulty in solving the predictive model in MPC/NMPC applications. However, applying constant design parameters cannot maintain consistent control effects in all states. Meanwhile, the designed system relies too much on sensor-measured data, and thus it is difficult to thoroughly validate the safety of the system because of its high complexity. This means that any potential hardware/software faults will endanger the engine. Therefore, this paper first presents a novel nonlinear mapping relationship to adaptively tune the tracking weight online with the change of Power Lever Angle (PLA) and real-time relative tracking error. Thus, without introducing additional design parameters, an Adaptive Tracking Weight-based SIMPC (ATW-SIMPC) controller is designed to improve the control performance in all operating states effectively. Then, a Primary/Backup Hybrid Control (PBHC) strategy with the ATW-SIMPC controller as the primary system and the traditional speed (Nf) controller as the backup system is proposed to ensure safety. The designed affiliated switching controller and the real-time monitor therein can be used to realize reasonable and smooth switching between primary/backup systems, so as to avoid bump transition. The PBHC system switches to the Nf controller when the ATW-SIMPC controller is wrong because of potential hardware/software faults; otherwise, the ATW-SIMPC controller keeps acting on the engine. The main results prove that the ATW-SIMPC controller with the optimal nonlinear mapping relationship, compared with the existing SIMPC controller, uplifts the dynamic control performance by 32% and reduces overshoots to an allowable limit, resulting in a better control effect in full state. The comparison results consistently indicate that the PBHC can guarantee engine safety in occurrence of hardware/software faults, such as sensor/onboard adaptive model faults. The approach proposed is applicable to the design of a model-based engine intelligent control system.

Fast reconstruction of EEG signal compression sensing based on deep learning

When traditional EEG signals are collected based on the Nyquist theorem, long-time recordings of EEG signals will produce a large amount of data. At the same time, limited bandwidth, end-to-end delay, and memory space will bring great pressure on the effective transmission of data. The birth of compressed sensing alleviates this transmission pressure. However, using an iterative compressed sensing reconstruction algorithm for EEG signal reconstruction faces complex calculation problems and slow data processing speed, limiting the application of compressed sensing in EEG signal rapid monitoring systems. As such, this paper presents a non-iterative and fast algorithm for reconstructing EEG signals using compressed sensing and deep learning techniques. This algorithm uses the improved residual network model, extracts the feature information of the EEG signal by one-dimensional dilated convolution, directly learns the nonlinear mapping relationship between the measured value and the original signal, and can quickly and accurately reconstruct the EEG signal. The method proposed in this paper has been verified by simulation on the open BCI contest dataset. Overall, it is proved that the proposed method has higher reconstruction accuracy and faster reconstruction speed than the traditional CS reconstruction algorithm and the existing deep learning reconstruction algorithm. In addition, it can realize the rapid reconstruction of EEG signals.

Vehicle Driving Safety of Underground Interchanges Using a Driving Simulator and Data Mining Analysis

In the process of driving in an underground interchange, drivers are faced with many challenges, such as being in a closed space, visual changes alternating between light and dark conditions, complex road conditions in the confluence section, and dense signage, which directly affect the safety and comfort of drivers in an underground interchange. Thus, driving simulation, building information modeling (BIM), and data mining were used to analyze the impact of underground interchange safety facilities on driving safety and comfort. Acceleration disturbance and steering wheel comfort loss values were used to assist the comfort analysis. The CART algorithm, classification decision trees, and neural networks were used for data mining, which uses a dichotomous recursive partitioning technique where multiple layers of neurons are superimposed to fit and replace very complex nonlinear mapping relationships. Ten different scenarios were designed for comparison. Multiple linear regression combined with ANOVA was used to calculate the significance of the control variables for each scenario on the evaluation index. The results show that appropriately reducing the length of the deceleration section can improve driving comfort, setting reasonable reminder signs at the merge junction can improve driving safety, and an appropriate wall color can reduce speed oscillation. This study indicates that the placement of traffic safety facilities significantly influences the safety and comfort of driving in underground interchanges. This study may provide support for the optimization of the design of underground interchange construction and internal traffic safety facilities.

Can deep learning compensate for sparse shots in the imaging domain? A potential alternative for reducing the acquisition cost of seismic data

Dense shots can improve the fold of subsurface imaging points, which is essential for the resolution of imaging results. However, dense shots significantly increase the cost of data acquisition, which is one of the major bottlenecks faced by seismic exploration. To address this issue, we speculate whether it is possible to construct an effective method to optimize the image made by stacking sparse shots and then generate an imaging result similar to the image made by stacking dense shots. In other words, we explore the possibility of using an optimization method to replace the dense shots in migration imaging, which would likely reduce the acquisition cost of seismic data. Deep learning can establish a nonlinear and complex mapping relationship by using data-driven strategies. Inspired by this, we use the convolutional neural network to establish a novel mapping relationship from the sparse-shot image to the dense-shot image by constructing a suitable training data set and designing a self-guided attention network architecture. We refer to this mapping relationship as shot compensation. We use the 2D Sigsbee2b model and the 3D SEG advanced modeling model to demonstrate the potential application of shot compensation in reducing the acquisition cost of seismic data. Moreover, a real 2D marine seismic data set is used to evaluate the effectiveness of shot compensation. Experimental results on synthetic and real data indicate that this shot compensation method can improve the quality of sparse-shot images and that the improved imaging results are similar to their corresponding dense-shot images.

Pseudo-medical image-guided technology based on 'CBCT-only' mode in esophageal cancer radiotherapy

Purpose To minimize the various errors introduced by image-guided radiotherapy (IGRT) in the application of esophageal cancer treatment, this study proposes a novel technique based on the 'CBCT-only' mode of pseudo-medical image guidance.Methods The framework of this technology consists of two pseudo-medical image synthesis models in the CBCT→CT and the CT→PET direction. The former utilizes a dual-domain parallel deep learning model called AWM-PNet, which incorporates attention waning mechanisms. This model effectively suppresses artifacts in CBCT images in both the sinogram and spatial domains while efficiently capturing important image features and contextual information. The latter leverages tumor location and shape information provided by clinical experts. It introduces a PRAM-GAN model based on a prior region aware mechanism to establish a non-linear mapping relationship between CT and PET image domains. As a result, it enables the generation of pseudo-PET images that meet the clinical requirements for radiotherapy.Results The NRMSE and multi-scale SSIM (MS-SSIM) were utilized to evaluate the test set, and the results were presented as median values with lower quartile and upper quartile ranges. For the AWM-PNet model, the NRMSE and MS-SSIM values were 0.0218 (0.0143, 0.0255) and 0.9325 (0.9141, 0.9410), respectively. The PRAM-GAN model produced NRMSE and MS-SSIM values of 0.0404 (0.0356, 0.0476) and 0.9154 (0.8971, 0.9294), respectively. Statistical analysis revealed significant differences (p < 0.05) between these models and others. The numerical results of dose metrics, including D98 %, Dmean, and D2 %, validated the accuracy of HU values in the pseudo-CT images synthesized by the AWM-PNet. Furthermore, the Dice coefficient results confirmed statistically significant differences (p < 0.05) in GTV delineation between the pseudo-PET images synthesized using the PRAM-GAN model and other compared methods.Conclusion The AWM-PNet and PRAM-GAN models have the capability to generate accurate pseudo-CT and pseudo-PET images, respectively. The pseudo-image-guided technique based on the 'CBCT-only' mode shows promising prospects for application in esophageal cancer radiotherapy.

A Hybrid Model for Long-Term Wind Power Forecasting Utilizing NWP Subsequence Correction and Multi-Scale Deep Learning Regression Methods

The accuracy of long-term wind power forecasting (WPF) is crucial for the efficient operation of grid systems. However, wind power generation is highly stochastic and intermittent due to the influence of weather, which makes long-term WPF less effective. Numerical weather prediction (NWP) data contains valuable weather forecast information, which can mitigate the negative effects of stochastic weather fluctuations on WPF. However, the accuracy of NWP data decreases over time, and multiple NWP data can have redundancy and errors that make it challenging to extract valid information. Reducing the variety and errors in NWP data and using more effective information extraction methods are essential for improving long-term WPF performance. In this article, we propose a novel long-term WPF hybrid model that corrects NWP wind speed and uses multi-scale deep learning regression prediction to exclude excessive NWP data. We use only the corrected NWP wind speed data to establish a nonlinear mapping relationship with actual power data. The validation case study shows that our proposed model reduces the mean squared error (MSE) and mean absolute error (MAE) by 65.0% and 43.8%, respectively, compared to the current state-of-the-art time series forecasting model in a seven-day forecasting scenario.

A Semi-Supervised Matrixized Graph Embedding Machine for Roller Bearing Fault Diagnosis Under Few-Labeled Samples

Exploring historical measurement data-driven health monitoring schemes for roller bearings is a current research hotspot. In engineering practice, the type of fault data obtained is often unknown and requires expensive costs to be annotated. However, most current intelligent diagnostic methods are based on the assumption that the labeled fault data is sufficient, so as to effectively establish the nonlinear mapping relationship between monitoring signals and health status. For this issue, a newly intelligent diagnosis method based on semi-supervised matrixized graph embedding machine (SMGEM) is proposed. In SMGEM, the geometric similarity relationship of unlabeled and labeled samples is obtained, which is subsequently embedded by incorporating a manifold regularization into SMGEM model, so that SMGEM can use the structure information of unlabeled samples to assist modeling. Meanwhile, a weighted nuclear norm (WNN) is used to highlight the importance of large singular values, so that a more accurate weight matrix can be constructed. The proposed method is verified by several roller bearing fault datasets, and experimental results demonstrate that the proposed semi-supervised diagnosis method can use few labeled samples to obtain a better identification accuracy.

Construction risk analysis of cantilever casting mid-span closure of bridge in reservoir area

In order to assess the impact of frequent microseismic and cyclic water flow forces on the bridge cantilever casting construction in the Three Gorges reservoir area, a limited element-quantum-orthogonal experimental design-neural network-Monte Carlo simulation method was used to analyze the risk assessment of the construction. The orthogonal experimental design method is utilized to generate probability distribution samples of water velocity, water height, and microseismic acceleration. These samples are then combined with a finite element model and trained using a BP neural network. This approach aims to establish a nonlinear mapping relationship between each control parameter and the axial deviation value in the mid-span closure, and then the Monte Carlo simulation is used to generate the random influence parameter, which can be substituted into the trained BP neural network to predict the risk of the error in the mid-axial deviation. The results suggested that the risk probability of the medial axis deviation of the closure section due to the coupling effect of water flow force and high-frequency microseismicity had reached 5.06%, and the risk of line deviation of the cantilever casting construction of the bridge in the reservoir area needed to be paid sufficient attention and prevented. The risk of line deviation of the cantilever casting construction of the reservoir bridge needs to attract enough attention and precaution.

An Accurate Torque Control Strategy for Permanent Magnet Synchronous Motors Based on a Multi-Closed-Loop Regulation Design

Torque control accuracy is a significant index of permanent magnet synchronous motors (PMSMs) and affects the safety of many applications greatly. Due to the strong nonlinearity of the motor as well as the disturbance of non-ideal factors such as temperature fluctuation and the parameter error in field-oriented control (FOC), it is undoubtedly difficult to accurately control the actual output torque. Meanwhile, the parameter differences between motors and sensors during mass production and the assembly process affect the consistency of output torque and even increase the factory failure rate of the motor. No torque sensor is implemented due to the cost and limited space. Accurate estimation of the motor torque becomes essential to realize the closed-loop feedback for torque and improve the accuracy at a lower cost. In this paper, a look-up table (LUT) model that can reflect the nonlinear mapping relationship between power and torque is established based on numerous offline experiments, which avoids the calculation of complex losses. A multi-closed-loop control strategy is proposed to dynamically adjust the amplitude and angle of the preset current command, respectively, to improve the torque accuracy. The effectiveness of the strategy has been validated by experimental results.

Multi-Objective Optimization of Injection Molding Process Parameters for Moderately Thick Plane Lens Based on PSO-BPNN, OMOPSO, and TOPSIS

Injection molding (IM) is an ideal technique for the low-cost mass production of moderately thick plane lenses (MTPLs). However, the optical performance of injection molded MTPL is seriously degraded by the warpage and sink marks induced during the molding process with complex historical thermal field changes. Thus, it is essential that the processing parameters utilized in the molding process are properly assigned. And the challenges are further compounded when considering the MTPL molding energy consumption. This paper presents a set of procedures for the optimization of injection molding process parameters, with warpage, sink marks reflecting the optical performance, and clamping force reflecting the molding energy consumption as the optimization objectives. First, the orthogonal experiment was carried out with the Taguchi method, and the S/N response shows that these three objectives cannot reach the optimal values simultaneously. Second, considering the experimental data scale, the back propagation neural network updated by the particle swarm optimization method (PSO-BPNN) was applied to establish the complex nonlinear mapping relationship between the process parameters and these three trade-off objectives respectively. Then, the Pareto optimal frontier of the multi-objective optimization problem was attained by multi-objective particle swarm optimization using a mutation operator and dominance coefficient algorithm (OMOPSO). And the competitive relationship between these objectives was further confirmed by the corresponding pairwise Pareto frontiers. Additionally, the TOPSIS method with equal weights was employed to achieve the best optimal solution from the Pareto optimal frontier. The simulation results yielded that the maximum values of warpage, sink marks, and clamping force could be reduced by 7.44%, 40.56%, and 5.56%, respectively, after optimization. Finally, MTPL products were successfully fabricated.

A Shear Wave Prediction Based on Levenberg-Maequart Algorithm

Shear wave velocity plays an important role in geophysics. Usually, the compressional velocity information is contained in the log data, but the shear velocity information is missing due to the cost and imperfect technology. Therefore, accurate prediction of shear velocity is a necessary step in seismic inversion. In order to solve this problem, we propose a shear velocity prediction method based on Levenberg-Marquardt algorithm. It grows from the Elman neural network algorithm, which establishes a nonlinear mapping relationship between shear velocity and other log curves by optimizing the log curve weights, and achieves a reasonable prediction of shear velocity.

Application of Deep Learning in Rayleigh Dispersion Curves Inversion

The inversion of the Rayleigh wave dispersion curve is a fundamental step in determining the shear wave velocity profile. However, existing inversion methods, such as linear and nonlinear approaches, have inherent limitations. In order to overcome these challenges and enhance the speed and accuracy of the inversion, this paper proposes a novel deep learning-based approach. We establish a method to construct sample data that ensures ergodicity and evolutive orderliness of the near-surface velocity profile using a constrained Markov decision process to address these limitations. We then design a deep learning model to capture the nonlinear mapping relationship between the dispersion sequence data and velocity structure. Training the model on the generated sample data accurately predicts the subsurface velocity profile based on the dispersion curve. The proposed method is evaluated through theoretical model experiments, demonstrating high computational accuracy.

A modeling method of wide random forest multi-output soft sensor with attention mechanism for quality prediction of complex industrial processes

Complex industrial production processes often involve multiple product quality indicators that are interrelated. There exists a complex nonlinear mapping relationship between the operational input feature variables and multiple output target quality variables, making it difficult to accurately model through first-principle models. In order to fully capture the complex relationship between measurable variables and difficult-to-measure quality variables, and achieve accurate prediction of multiple output variables to meet the needs of practical industrial sites, this paper proposes a broad random forest-based multi-output soft sensor modeling method based on the idea of attention mechanism derived from the concept of broad learning systems. This method comprehensively considers the dynamic impact of different feature variables on the target quality indicators in actual production processes. The attention mechanism assists the soft sensor model in capturing contextual information better when dealing with long sequences, with a focus on the relevant parts related to the current task. Additionally, the interpretable random forest algorithm is employed as the weight estimator for the basic feature learning unit of Broad-based learning, enabling regression modeling of multiple target quality variables. The use of Broad-based random forest improves the model’s learning ability, interpretability, and generalization capability. To validate the reliability of the proposed method, it was applied to real industrial cases. The results demonstrated that the multi-output quality variable prediction performance of the proposed soft sensor outperforms existing soft sensors in terms of prediction accuracy. This indicates promising industrial application prospects for the proposed method.

An automated computation method for flexible protection systems based on neural networks

An automated computation method is proposed in this paper to address the problems of enormous computational difficulty and low efficiency in flexible protection systems. This method uses neural networks to demonstrate the complex nonlinear mapping relationships between the system’s specifications and performance. Use the finite element numerical simulation model validated by a full-scale impact test on the 750 kJ passive flexible protection system for massive parametric analysis, and extract the critical computation results as the label space. Conducted mechanical tests on critical components, analyzed the mechanical characteristics of each component, and extracted training features to form a sample space. The label space and sample space formed a dataset containing 1828 sets of data and used 1700 sets of data to train a neural network while using the remaining 128 sets of data to test. The maximum predicted error was 8.33 %, and the maximum average error was 3.73 % from the test. The proposed method can complete the computation in just 6 s, with a computation speed of 520 times that of the fastest finite element method, significantly improving the computation efficiency of the flexible protection system and reducing the computational difficulty.

Machine learning for carbonate formation drilling: Mud loss prediction using seismic attributes and mud loss records

Due to the complexity and variability of carbonate formation leakage zones, lost circulation prediction and control is one of the major challenges of carbonate drilling. It raises well-control risks and production expenses. This research utilizes the H oilfield as an example, employs seismic features to analyze mud loss prediction, and produces a complete set of pre-drilling mud loss prediction solutions. Firstly, 16 seismic attributes are calculated based on the post-stack seismic data, and the mud loss rate per unit footage is specified. The sample set is constructed by extracting each attribute from the seismic trace surrounding 15 typical wells, with a ratio of 8:2 between the training set and the test set. With the calibration results for mud loss rate per unit footage, the nonlinear mapping relationship between seismic attributes and mud loss rate per unit size is established using the mixed density network model. Then, the influence of the number of sub-Gausses and the uncertainty coefficient on the model's prediction is evaluated. Finally, the model is used in conjunction with downhole drilling conditions to assess the risk of mud loss in various layers and along the wellbore trajectory. The study demonstrates that the mean relative errors of the model for training data and test data are 6.9% and 7.5%, respectively, and that R2 is 90% and 88%, respectively, for training data and test data. The accuracy and efficacy of mud loss prediction may be greatly enhanced by combining 16 seismic attributes with the mud loss rate per unit footage and applying machine learning methods. The mud loss prediction model based on the MDN model can not only predict the mud loss rate but also objectively evaluate the prediction based on the quality of the data and the model.