60results about How to "Convergent stability" patented technology

The invention relates to the technical field of knowledge graph completion, in particular to a large model knowledge graph completion method and system based on causal guidance. The method comprises the following steps: acquiring a target knowledge graph and an input triple to be complemented, performing structured analysis on an input triple relationship, extracting key topological characteristics, constructing a structured mediation variable, mapping the structured mediation variable into a structure guide prefix, injecting the structure guide prefix into large model input, and constructing a double-path inference model to generate an inference prediction result; and meanwhile, a gradient sensing dynamic loss balance mechanism is introduced, the loss weight is adaptively adjusted according to reasoning feedback, and finally a more accurate and stable knowledge graph completion result is output. According to the method, the controllability, interpretability and training stability of the reasoning process can be enhanced while the knowledge graph completion precision is improved.

The invention provides a multi-agent strategy network training method and device, a strategy generation method and device, equipment and a medium. The multi-agent strategy network training method comprises the steps that local observation states and global observation states of all agents and rewards for executing action strategies in the execution process of the action strategies generated based on a currently updated strategy network are collected; determining a target obstacle function value corresponding to each safety evaluation network based on the local observation state and a pre-constructed random discrete diagram control barrier function; determining the security constraint condition of each agent based on the approximate value of the obstacle function output by each security evaluation network; determining an advantage value of each agent action based on a state value prediction value output by the task value evaluation network and the collected rewards; and under the constraint of a security constraint condition, guiding the strategy network to perform parameter updating by taking the advantage value as an updating direction signal. According to the method, the training oscillation or divergence phenomenon caused by safety signal distortion is reduced, and a smoother and more stable convergence process is realized.

This application discloses a method and apparatus for pre-training large-scale models based on Bayesian domain reweighting. The method includes: dividing the pre-training corpus into multiple data domains and obtaining corresponding empirical loss signals; learning gamma prior hyperparameters based on the empirical loss signals using a prior prediction network to determine the morphological constraints of the variational posterior distribution; optimizing the morphologically constrained variational posterior distribution based on the optimization objective to obtain a converged domain weight distribution, from which the final domain weights are determined, and using this to perform weighted sampling of the pre-training corpus, achieving mixed sampling of pre-training data for large language models. This solves the problem in existing large-scale model pre-training processes where the learning trajectory of domain weights is unstable, making it impossible to stably and accurately learn the optimal domain weights while simultaneously achieving efficient computation, resulting in low training efficiency and weak generalization ability of large models. This method achieves efficient utilization of multi-domain data and improves the comprehensive performance of large models on different tasks.

The application discloses an emotion recognition method based on online cross-modal knowledge distillation, comprising the following steps: acquiring electroencephalogram and electrocardiogram original signals and windowing and cutting; constructing electroencephalogram and electrocardiogram student models, extracting intermediate features from each modal data segment through an encoder, and obtaining non-normalized prediction output through a classifier; constructing a teacher probability distribution through a joint encoder fusion; introducing adaptive contrast loss to align the cross-modal intermediate features, introducing distillation loss to constrain the prediction probability distribution of each modal to align with the teacher probability distribution; synchronously optimizing new student model parameters through online collaborative training; and performing actual inference prediction based on the student model after training. The application combines double modal signals to make up for the defects of single modal information, excavates the complementarity of modes, realizes dynamic generation of teacher supervision signals and real-time collaborative learning of modes through online distillation, does not increase test calculation overhead, effectively improves the recognition accuracy, model robustness and generalization ability, and has good application prospect.

This invention discloses a multi-objective optimization method for flexible distribution systems based on reward decoupling normalization reinforcement learning. This method constructs a multi-objective deep reinforcement learning model based on group reward decoupling normalization strategy optimization to perform multi-objective optimization scheduling for complex constraints in flexible distribution systems, such as renewable energy fluctuations, load changes, and scheduling resource coordination. The method designs the three objectives—minimizing operating costs, minimizing voltage offset, and minimizing photovoltaic curtailment—as independent reward functions, guiding the flexible distribution system's operation strategy to focus on multiple dimensions, including economy, environmental friendliness, and safety. The group reward decoupling normalization strategy optimization algorithm independently normalizes each reward, avoiding the collapse problem of multiple reward signals in advantage estimation and effectively preserving the relative differences between the objectives.

The present application belongs to the technical field of Internet of Vehicles and communication security, and discloses a federated learning training method based on vehicle state perception and applicable to dynamic Internet of Vehicles. The method collects vehicle running state information of vehicle clients, preliminarily screens candidate vehicle clients, calculates comprehensive scores of the screened vehicle clients, adopts a client selection strategy combining selection of the top-ranked clients with random supplementation to determine a set of vehicle clients participating in the current round of federated learning, configures local training parameters for the selected vehicle clients according to vehicle speed and computing capacity, performs local training on local data sets to obtain model updates, performs weighted aggregation on the model updates of the effective participating vehicle clients at the server end, introduces a momentum mechanism to smooth the model update process, and obtains new global model parameters. This method effectively improves the stability, robustness and training efficiency of federated learning in a high-dynamic Internet of Vehicles environment.

The application belongs to the technical field of cloud computing and artificial intelligence, and is used to solve the problem that the prior art cannot perceive the anchor point domination relationship within a batch and dynamically constrain the batch boundary accordingly. Specifically, the application is a cloud server data security management system, which comprises a task acquisition module, a conflict analysis module, a conflict relationship construction module, a grouping management module and an execution control module. The application constructs the potential domination strength based on a sliding window and the output sensitivity based on the model sensitivity and the permission level for each task, calculates the conflict coefficient for batch constraint, so that when the input feature value of a task exceeds three times of the quartile range of its own historical distribution, the conflict coefficient between the task and the high sensitivity task automatically exceeds the dynamic threshold, thereby the task is split into different batch groups. The task pair with the conflict coefficient exceeding the threshold is marked as a conflict and prohibited from being in the same batch, and the low permission task cannot enter the same batch as the high sensitivity task.

The application provides an industrial part point cloud denoising method based on multi-expert fusion and segmented iteration, which comprises the following steps: obtaining a point cloud data set; constructing a denoising model; a geometric perception multi-expert module takes the preprocessed noisy industrial part point cloud as input, and respectively extracts local learning features and local geometric features representing the local structure of the point cloud; a routing gate and a geometric perception multi-expert fusion mechanism are used to process the local learning features and the local geometric features to obtain fusion features; a segmented iterative updating module takes the fusion features as input, and obtains denoised point cloud data through a noise perception segmented iterative updating mechanism; the denoising model is trained by using the point cloud data set, a detail discriminator and a smooth discriminator are introduced to construct a double-discriminator constraint loss during the training process, and a reconstruction loss, a geometric consistency constraint loss, an optimal transport distribution constraint loss and a trajectory consistency constraint loss are introduced; and the trained denoising model is used for point cloud denoising. The application solves the problems of poor noise adaptability and loss of geometric features.

This invention relates to the field of microfluidic control technology and discloses a method and system for detecting and controlling microfluidic droplet formation. The method first constructs an ideal concentric ring geometric model by acquiring the target parameters of the microcapsule and establishes a force-shape coupling response matrix. Based on this, a preset velocity triplet is solved to generate the initial droplet. Then, droplet image sequences are acquired and fitted with real-time concentric ring groups. These are geometrically superimposed with the ideal concentric ring geometric model to generate a deviation ring domain, which is then decomposed hierarchically to obtain geometric deformation characteristic components. Finally, a coordinated velocity adjustment vector is calculated using the force-shape coupling response matrix, a composite fluid correction field is constructed, and the deviation ring ablation cycle is iteratively executed until geometric convergence is achieved. This invention effectively solves the oscillation problem caused by strong parameter coupling in multiphase flow fields by constructing a quantitative mapping and inverse decoupling mechanism between fluid and geometric shape, achieving coordinated and precise control of the microcapsule's external dimensions and internal structure.

This invention provides a method and system for testing the electromagnetic radiation immunity of components, relating to the field of electromagnetic immunity testing technology. The invention sets a range for the electromagnetic frequency and amplitude of the interference signal, generating multiple particles characterizing the interference signal within this range. Through iterative particle swarm analysis, it rapidly collects highly sensitive sample points that cause functional abnormalities in components during electromagnetic interference testing. By clustering these highly sensitive samples, it locates the electromagnetic interference sensitive range of the target component. Compared to full-band scanning and fixed-point scanning methods for electromagnetic interference testing, this invention enables automated and quantitative extraction of the electromagnetic sensitivity range, reducing subjective bias from manual location and improving analytical consistency and repeatability. Furthermore, by focusing on highly sensitive frequency bands and field strength ranges, it optimizes the test path, reduces redundant scanning points, and significantly improves the efficiency of electromagnetic interference testing.

This invention discloses a method and system for geological hazard identification based on multimodal semi-supervised learning. The method includes the following steps: acquiring labeled and unlabeled orthophotos and elevation models of geological hazard areas; constructing teacher and student networks containing dual-branch encoders, extracting features from the orthophotos and elevation models respectively, and linearly fusing them, generating ensemble predictions using learnable dynamic weights; in semi-supervised training, calculating the uncertainty of the student network's prediction results in real time, and triggering an exponential moving average update of the teacher network using the student network parameters only when the decrease in uncertainty compared to the historical mean exceeds a preset evolution threshold; optimizing the student network parameters by combining supervised loss, confidence-weighted consistency loss, and student historical consistency loss; inputting the test data into the trained student network, and outputting fine-grained geological hazard segmentation results. This invention can achieve high-precision automatic identification of geological hazards under small sample conditions.

The invention discloses an emotion recognition method based on online cross-modal knowledge distillation. The emotion recognition method comprises the steps that electroencephalogram and electrocardio original signals are obtained and subjected to window segmentation; constructing an electroencephalogram and electrocardio student model, extracting intermediate features from each modal data segment through an encoder, and obtaining non-normalized prediction output through a classifier; teacher probability distribution is constructed through joint encoder fusion; self-adaptive comparison loss alignment cross-modal intermediate features are introduced, and distillation loss is introduced to constrain prediction probability distribution of each modal to align to teacher probability distribution; new student model parameters are optimized synchronously through online cooperative training; and performing actual reasoning prediction based on a student model after training. According to the method, bimodal signals are combined to make up single-modal information defects, modal complementarity is mined, teacher supervision signal dynamic generation and modal real-time collaborative learning are realized through online distillation, test calculation overhead is not increased, identification precision, model robustness and generalization ability are effectively improved, and the application prospect is good.

The invention provides a fine-grained image classification model training method and system based on heterogeneous resolution, and relates to the technical field of model training. Obtaining a multi-scale image based on the high-resolution fine-grained image sample; performing mosaic processing on the multi-scale image to obtain a heterogeneous image sample simulating degradation; the images are input into a fine-grained image classification model, high-resolution fine-grained image samples are input into a standard flow network to obtain standard features, and heterogeneous image samples are input into an adaptive flow network to obtain adaptive features; generating an information density map based on the energy difference between the standard features and the adaptive features, taking the information density map as a weight to construct feature repair loss, and guiding the adaptive features to be aligned to the standard features; respectively mapping the adaptive features and the standard features to a semantic manifold space, and constraining the consistency of the adaptive features and the standard features in the semantic direction; and the model parameters are optimized in combination with the loss function to obtain a trained fine-grained image classification model, so that the robustness of the model and the accuracy of application to image classification are improved.

The invention relates to the technical field of medium-high frequency vibration control, and discloses a structure medium-high frequency vibration variable weight adaptive robust compensation type active control method, which comprises the following steps of: respectively inputting error signals into a minimum mean square controller, a robust controller and a weight controller for operation; a weight controller compares the squared error signal with a reconstructed reference signal mean value, and adaptively adjusts the weights of a minimum mean square controller and a robust controller according to the signal mean value; the first output control quantity and the second output control quantity are superposed and then act on a structural vibration control point through an electromagnetic actuator to obtain an output signal, the output signal is superposed with an original reference signal, and a new error signal is formed to serve as an input signal of a next round of control operation. According to the method, a reference signal reconstruction technology is adopted, a robust compensation mechanism is additionally arranged, an adaptive variable weight strategy and a convergence time and error robustness statistical method are configured, and the medium-high frequency vibration control effect of structures such as a new energy automobile electric drive assembly speed reducer shell is effectively improved.

The invention discloses a rock slag feature recognition method based on multi-sensor fusion. The method comprises the steps that 1, an industrial camera and a laser radar synchronously collect rock slag images and point cloud data; 2, an improved YOLO11n-seg network is constructed; 3, performing feature extraction on the rock slag image through an improved YOLO11n-seg network; 4, constructing total loss; 5, training the improved YOLO11n-seg network on the basis of the total loss; 6, segmenting a subsequent rock slag image to be processed to obtain a rock slag area segmentation map; 7, projecting the original point cloud data to obtain the point cloud of each rock slag area; and obtaining the rock slag volume based on the point cloud of each rock slag area. The method is simple in step, instance segmentation of the rock slag target area is achieved through the improved YOLO11n-seg network, the contour shape and the volume of the rock slag are obtained, and therefore a basis is provided for parameter optimization of follow-up TBM tunneling.

This invention discloses a reinforcement learning simulation step size control method based on closed-loop adaptive noise injection and entropy increase optimization strategy, belonging to integrated circuit computer-aided design technology. Specifically, the method involves: first, inputting the circuit netlist and interacting with the simulator using file read / write; next, establishing two new network output layers to convert the deterministic actions of the policy network's output into a probability distribution; then, obtaining the entropy regularization term based on the probability density function and weight coefficients, and adding it to the policy loss function; next, generating Gaussian-distributed exploration noise and adjusting the standard deviation of the noise at each step using a PID controller; finally, adjusting the policy network parameters using a gradient update method and injecting the exploration noise into the output actions to obtain the final time step. Using this invention helps enhance the simulation stability of pseudo-transient analysis, improves simulation efficiency, and provides a new method for DC analysis.

The present application relates to a kind of nuclear power unit one loop boron concentration control system and method.The lower outlet of the nuclear power unit one loop in the system is connected with the first end of electric regulating valve and the first end of hydrogenation station respectively through pipeline, the second end of electric regulating valve is connected with the input end of volume control box through pipeline, the output end of volume control box is connected with the first end of filling pump through pipeline, the second end of hydrogenation station is connected with the first end of filling pump through pipeline.The second end of filling pump is connected with the recovery port of one loop.Electric regulating valve is connected with control terminal, and control terminal is used to adjust the opening of electric regulating valve.The present application can realize the rapid stable convergence of one loop boron concentration by the flexible adjustment of the proportion of RCV system lower flow through volume control box according to various working conditions.

The application discloses an image coding, decoding and compressing method and system based on vector quantization entropy modeling, which comprises the following steps: obtaining a first feature map of an image to be coded; obtaining a super-prior binary code stream and super-prior information of the first feature map; quantizing the first feature map by using a lattice vector quantizer; projecting the quantized first feature map to an integer coefficient domain of the lattice vector quantizer to obtain a second feature map; modeling the distribution of the second feature map as a Gaussian distribution independent of each dimension, and predicting the mean and variance of the Gaussian distribution by combining the super-prior information and a spatial context model; dividing the integer coefficient domain of the lattice vector quantizer by using a relaxed boundary, and performing probability estimation and arithmetic coding according to the predicted mean and variance of the Gaussian distribution to obtain a feature binary code stream; and combining the feature binary code stream with the super-prior feature binary code stream to obtain a compressed image binary code stream. The application improves the rate-distortion performance of image compression by using a more efficient lattice vector quantization on three-dimensional features.

The present application relates to the technical field of data processing, and particularly relates to a parameter optimization method for yucca saponin extraction process based on data processing, which comprises the following steps: obtaining multi-dimensional parameter data in the yucca saponin extraction process to construct all process parameter combinations, and initializing a particle swarm containing multiple particles, and each particle represents a set of process parameter combinations; outputting the predicted yield of the process parameter combination in the target iteration by using a machine learning model, and taking the particle corresponding to the process parameter combination with the highest predicted yield as the optimal particle in the target iteration; updating the particle swarm by using an improved particle swarm algorithm to output a global optimal particle, and taking the process parameter combination corresponding to the global optimal particle as the best extraction process parameter. The present application solves the problem of lack of flexible adjustment ability of the existing algorithm in different iteration stages.

The application provides a multi-person AR system data offloading and resource allocation method under a 6G network framework, relates to the technical field of 6G mobile communication, selects an AR user as a Host user in each time slot, and the rest are Resolver users; the interaction weight between the Host user and the Resolver user is calculated by using a graph attention network; a data offloading model based on Stackelberg game is established, a multi-party utility function is constructed, a double-layer optimization problem is converted into a single-layer optimization problem, and a sequential quadratic programming algorithm is adopted to solve, so that the optimal data offloading decision is obtained; the resource allocation process is constructed as a Markov decision process, a reward function containing average delay and delay variance is defined, a double-delay deep deterministic policy gradient algorithm is adopted to output a resource allocation strategy of a continuous action space; through the combination of game theory and reinforcement learning, the environmental data redundancy and server rendering load are effectively reduced, the interaction consistency among multi-users is optimized while ensuring low delay, and the collaborative experience of the multi-person AR system is improved.

This invention discloses a robust aggregation method and system for federated learning based on multidimensional logs and deep reinforcement learning, applicable to federated learning systems including a server and multiple clients. The method includes: initializing a global model, a deep reinforcement learning agent, and an experience replay pool on the server side; sending the global model to each client and receiving local model updates; extracting multidimensional feature vectors and constructing a feature matrix through standardization and normalization; calculating the statistical moments of all client feature vectors, constructing a global state vector independent of the number of clients, inputting it into the agent's policy network to obtain an indicator weight vector; obtaining a corresponding comprehensive reputation score from the indicator weight vector and the feature vector of each client, introducing a temperature coefficient and normalizing it to obtain aggregation weights; weighted aggregation of the local model updates of all clients to update the global model; and updating the agent network parameters based on the accuracy of the new global model on the validation set as a reward signal.

The present application relates to a kind of unmanned airborne radar jamming intention recognition method, belong to electric digital data processing field, method includes: signal preprocessing and frequency domain adaptive enhancement;Multi-scale feature extraction and priori guide Transform coding;Dual granularity joint classification identification;Timing correlation analysis and structured semantic generation;Large language model collaborative reasoning and strategy output.The present application provides low SNR scene recognition capability: through frequency domain adaptive weighting mechanism, key frequency band can be automatically strengthened and background noise is inhibited, therefore, in the ground SNR scene, recognition rate is significantly improved, weak interference detection capability is enhanced.Improve the identification performance of composite interference field: through multi-scale convolution fusion, short pulse and long suppression interference are simultaneously identified, improve complex scene coverage capability, compared with single scale structure, composite interference recognition accuracy is significantly improved, and model generalization ability is enhanced.

The application discloses an AoI-aware incentive method for railway data sharing under privacy constraints, and comprises the following steps: a railway data sharing model is constructed, including a data requester and multiple data providers; a utility maximization problem is constructed with the aim of maximizing the self-utility of the data requester and the data providers, including an incentive payment strategy and a data uploading strategy; the utility maximization problem is modeled as a multi-agent Markov decision process; the optimal strategy of each agent is learned by using IMASAC, the multi-agent Markov decision process is solved, and the optimal incentive payment of the data requester and the optimal data uploading amount of the data providers are obtained. The application can depict freshness-aware double-layer requester-provider interaction and avoid the privacy bottleneck caused by centralized critics.

The application belongs to the technical field of recommendation systems, and particularly relates to a hierarchical latent inference recommendation based on a large language model and a model training method. A category level and an item identification level hierarchical representation are constructed for items, a recommendation model is built based on a large language model, the recommendation model performs hierarchical latent inference according to a user historical interaction sequence, generates a category level latent state, and then generates a user collaborative filtering vector matching an item identification level representation; a phased progressive training strategy is adopted, and real category level representations and model self-generated category level latent state supervision vectors are used to generate vectors respectively. In the recommendation stage, the user historical interaction sequence is input into the trained model, the hierarchical latent inference is performed to output the user collaborative filtering vector, similarity calculation is performed between the user collaborative filtering vector and an item library identification level collaborative representation, and recommended items are determined.

The invention relates to a data-driven load frequency control method for a power system containing a network-forming converter. The method comprises the following steps: firstly, according to a power system topology, constructing a time-delay nonlinear load frequency control model containing a network construction type converter; taking the model as an environment, simulating operation acquisition data by using a PID controller, and constructing a pre-training data set; then, an agent of an actor-commentator framework is constructed, and offline pre-training is completed through a data set; and finally, accessing the intelligent agent into a system for online interaction, outputting a frequency modulation instruction by an action network, and iteratively updating parameters by combining an award and a loss function to realize self-adaptive control. Compared with the prior art, the method has the advantages of being adaptive to the dynamic characteristics of the GFM, efficient, high in robustness and the like.

The application discloses a method and system for adjusting the feel of an electrocoagulation surgical instrument. The method includes: performing exponential moving average incremental updating based on historical pressure thresholds and robust pressure samples to obtain incrementally updated pressure samples; performing online statistics on the incrementally updated pressure samples to obtain a thermal center position coordinate, a holding pressure mean value, a holding pressure standard deviation, and a pressure rise rate; obtaining a reference threshold based on the holding pressure mean value and the holding pressure standard deviation, the reference threshold including a normal holding pressure interval and an overpressure threshold; and obtaining an operation trigger confirmation duration based on the pressure rise rate. The application learns individual holding habits online and derives the reference threshold and the confirmation duration, so that the instrument actively adapts to the operator, presents a fit feel, shortens the learning curve, reduces unnatural holding postures and hand fatigue, realizes online calibration with automatic convergence, is resistant to environmental drift, and ensures stable convergence of the calibration without divergence.

The invention relates to a transient detection method for the thickness of deposited ash. The method comprises the following steps: firstly, acquiring actually measured temperatures of a first thermocouple arranged on the inner side of a pipe wall and a second thermocouple arranged in the pipe wall and a flue gas temperature; on the basis of the temperatures, thermophysical parameters under the current working condition are calculated, an assumed ash deposition thickness is initialized, and then the convective heat transfer coefficient between the flue gas and the ash layer is solved. Then, a transient temperature field control equation from the pipe wall to the outer surface of the ash layer is established in combination with the assumed thickness and the heat exchange coefficient; and iteratively solving the equation through an explicit central difference method to obtain the simulated temperature of the second thermocouple position. And finally, comparing the simulated temperature with an actual measurement value of the second thermocouple: if an error is smaller than a set threshold value, outputting the current assumed thickness as a detection result; otherwise, according to the condition that the simulation value is higher than or lower than the measured value, the assumed ash deposition thickness is correspondingly increased or decreased, and the calculation process is repeated until convergence is achieved. According to the invention, starting from unstable heat conduction, the heat conduction process under variable working conditions can be simulated more truly, and transient and accurate online detection of the soot deposition thickness is realized.

The invention relates to an image processing method and device, computer equipment, a storage medium and a program product. The method comprises the following steps: acquiring an image acquired by a robot; performing semantic segmentation and mask generation on the image to obtain a task area, a non-task area and mask image data of the non-task area; performing disturbance processing on the non-task area to obtain a disturbed non-task area; and based on the task area, the post-disturbance non-task area and the mask image data, generating a target image for robot imitation learning. By adopting the method, the problem of distraction in the imitation learning process can be effectively avoided.

The present application relates to the technical field of adaptive parameter control processing, and specifically discloses an adaptive parameter control system for ozone deodorization, which continuously collects the ozone inlet concentration, ozone outlet concentration, oil phase feed flow, ozone inlet flow, pressure difference between the bottom and top of the reaction tower, and reaction temperature in the gas-liquid reaction tower; according to the collected data, the gas-liquid exchange activity value is estimated by alternately executing state recursion and observation correction; the activity estimation value is compared with the initial activity value to calculate the activity decay percentage, a feedforward compensation enable signal is generated when the threshold value is reached, the correction coefficient is calculated according to the current activity estimation value, target dissolution efficiency and current dissolution efficiency, the corrected ozone inlet flow set value is obtained and the flow is adjusted; the actual ozone inlet flow and newly collected ozone outlet concentration are fed back to the state estimation; the present application can avoid reverse regulation errors caused by mass transfer efficiency decay and stabilize the deodorization effect.

This invention relates to a method for diagnosing UAV motor faults based on dilated convolution residuals. The method comprises a data preprocessing module that converts one-dimensional vibration signals collected during UAV motor operation into two-dimensional grayscale images; a feature extraction module that extracts features from the input grayscale images to generate multi-channel feature maps; and a fault classification module that compresses the feature dimensions of the multi-channel feature maps, performs dimensionality mapping, and outputs the probabilities of various fault types, which constitute the fault diagnosis results. This invention improves fault identification accuracy by expanding the receptive field through dilated convolution and focusing on key features using a CBAM attention module, effectively capturing early, weak fault signals in the motor. The grayscale image conversion method in the data preprocessing stage does not require manual preset feature parameters, and the model can adaptively learn motor fault features under different operating conditions (such as different speeds and loads), with a generalization error of less than 5%.