245 results about "Sample Weight" patented technology

The invention provides a noise robust text-to-image pedestrian retrieval method and device based on a soft label. The method comprises the following steps: calculating cosine similarity based on image global features and text global features, and normalizing the cosine similarity to generate a soft label representing image-text pairing confidence; distributing a sample weight for each training sample according to the soft label, and obtaining a joint weight for current iteration in combination with a dynamic weight factor progressively increased along with the training process; respectively constructing cross-modal contrast learning loss and similarity distribution matching loss by using the joint weight, and carrying out weighted summation on the cross-modal contrast learning loss and the similarity distribution matching loss to obtain a total loss function; and updating parameters of the image encoder and the text encoder by using the total loss function until training convergence, and obtaining a cross-modal alignment model. According to the method, robust cross-modal alignment can be realized, and the pedestrian retrieval accuracy in a noise scene is improved.

The invention discloses a multi-agent medical inquiry system based on an improved GRPO algorithm, and belongs to the field of intelligent medical treatment. According to the method, firstly, a new medical data set containing illusion answers is constructed through an open source medical data set, clinical correlation of the data set is evaluated through multiple large models, and weights are given to samples; then, an improved GRPO algorithm is adopted to optimize the multi-modal large model, and the model is encouraged to explore different answers by progressively decreasing KL divergence constraints in the optimization process; the optimized model is applied to construction of a medical question and answer system, the system works cooperatively through multiple agents including a user agent, a hospital guide agent, a doctor agent and an expert agent, and automatic illness state analysis, consultation department judgment, medical opinion generation and specialized correction of an inquiry result are achieved. According to the invention, the accuracy and specialty of medical inquiry are effectively improved, the misdiagnosis rate is reduced, the medical efficiency is improved, and more accurate and comprehensive inquiry services are provided for users.

The invention relates to an embankment slope stability assessment method based on artificial intelligence, and belongs to the technical field of embankment slope monitoring and assessment. The method comprises the following steps: acquiring and marking embankment slope stress sensing data; dividing the data into a plurality of spatio-temporal data blocks; constructing a stability evaluation model; extracting multi-scale convolution features by adopting multi-scale cavity convolution of a high-frequency channel and pooling-deconvolution operation of a low-frequency channel to obtain a fused multi-scale feature matrix; a hidden state sequence is obtained through a space attention mechanism and a double-door-setting mechanism; calculating time interval saliency based on the hidden state vector, then calculating a weighted feature vector, further obtaining a weighted feature matrix, and processing through deep convolution and point-by-point convolution to obtain a pooling feature vector; carrying out stability evaluation grade classification through learnable category prototype and gating feature transformation; and dynamically adjusting sample weight and constraint attention distribution by adopting a total loss function. According to the method, the progressive instability identification capability can be improved.

The invention provides a data security encryption method and system based on high-security machine learning, and the method comprises the steps: collecting an access behavior of to-be-encrypted data, and calling context and environment meta-information to construct a structured feature sequence; establishing index mapping between the feature sequence and a historical encryption execution record, and extracting significant features by using a feature scoring and information gain evaluation method; the system executes multi-level decision path search based on hierarchical condition division according to the significant feature sequence of the current task, and judges the security level and the encryption risk situation of data; the system constructs an encryption control sequence including a parameter instruction block and an execution logic structure according to the task risk level and the strategy index; the encryption engine analyzes the instruction packet and executes a corresponding encryption task, and generates a complete record file containing metadata and a tracking tag at the same time; and the system performs statistical analysis on decryption feedback logs and encryption abnormal records, and optimizes the strategy dictionary by adopting a sample weight updating and feature re-evaluation method.

The invention discloses a hyperspectral image classification method based on frequency domain denoising and element gradient correction, and the method comprises the following steps: carrying out the preprocessing of all hyperspectral image data, and dividing an overall training sample set formed by the processed hyperspectral images into a training set and a verification set; constructing a sample weighting model based on frequency domain denoising and element gradient correction; in the training process, a parameterization frequency spectrum gating sensing transformation module is utilized to map features to a frequency domain through discrete Fourier transform, a learnable frequency spectrum response function is utilized to adaptively suppress spectrum jitter noise, and finally pure features are reconstructed. And automatically constructing a high-confidence pseudo-clean verification set based on a Gaussian mixture model and time domain consistency. According to the method, a time domain momentum updating mechanism is introduced, the variance of statistical estimation is effectively smoothed, random interference caused by training fluctuation is resisted, and the accuracy of pseudo clean set construction and the convergence stability of overall model training are further improved.

The invention discloses an efficient graph comparison learning method for social recommendation. As an emerging self-supervised learning normal form, graph contrast learning is excellent in response to data sparseness and cold start due to the fact that the graph contrast learning can effectively capture similarity and heterogeneity characteristics in a graph structure, although the learning normal form achieves a good effect in a recommendation system, the graph contrast learning can be used for solving the problems of data sparseness and cold start. However, the method still faces three defects: (1) average neighbor aggregation and a non-adaptive representation reading mechanism are adopted in a message propagation process, and high-quality node representation is difficult to learn; (2) a visual angle is enhanced by depending on a random disturbance generation graph during intervention of comparative learning, which may destroy the inherent structure of graph data and further weaken the accuracy of the model; and (3) equally treating all observation samples during parameter optimization, and neglecting the difference influence of positive samples in different training stages. Specifically, aiming at the problems, the invention provides an efficient graph contrast learning method (EGCL for short). The method comprises the following steps: firstly, designing a graph adaptive propagation module, improving an information propagation rule of a graph neural network by referring to a thermonuclear thought and an attention mechanism, and realizing differentiated aggregation of neighbor nodes by adopting a learnable weight distribution strategy; secondly, designing a double contrast learning normal form which does not need graph enhancement, and realizing mutual promotion of node characterization through intra-domain contrast learning (inter-CL) and inter-domain contrast learning (inter-CL); and finally, introducing a sample weight adaptive efficient optimization algorithm, converting the training process into a double-layer optimization problem, and adaptively adjusting the contribution degree of each sample to model optimization in different stages.

The invention discloses an uncertainty guided few-sample harmful speech detection method (U-GIFT). According to the method, a pre-training language model is finely adjusted based on a small number of labeled samples, and a semi-supervised self-training and uncertainty guiding strategy is combined. Monte Carlo Dropout is started in the reasoning stage, multiple times of random forward propagation are carried out to obtain sample posterior distribution, prediction entropy and information gain are calculated, pseudo-label samples are sorted and screened, and only high-confidence samples are selected to be added into a training set. And in order to reduce the influence of a pseudo labeling error, designing a stability weighting mechanism, giving a sample weight according to a prediction variance, and constructing a joint loss function, so that the model preferentially learns a stable sample to improve the detection performance. According to the method, the semantic and attention mechanism of the pre-training model is utilized, the detection effect is remarkably improved under the conditions of few samples, imbalance, multiple languages and cross domains, models such as BERT, RoBERTa, XLM-R, LLaMA2 and DeepSeek-R1 are compatible, and the method is suitable for content auditing and risk prevention and control.

The invention discloses a crop disease and insect pest detection algorithm fusing self-attention and a sample weighting mechanism. A YOLOv8 network is used as a basic model; through a two-branch self-attention DF-MSA architecture, the ability of the model to acquire pest target position information is enhanced; a cross-layer feature fusion module is introduced into the feature pyramid network structure, and multi-scale information is fused; and a sample weighting function is adopted to reduce the influence of difficult sample imbalance on a detection result, so that the detection accuracy is improved. And inputting the trained IP-YOLO by using a test set sample, and outputting a result. The method has high stability and robustness, and an effective means is provided for improving the performance of a crop disease and pest detection system.

The invention discloses a federated backdoor defense method based on decoupling contrast learning, and the method comprises the steps: training a backdoor model based on a backdoor sample, and immediately stopping training after the backdoor model converges on the backdoor sample; respectively extracting a penultimate layer vector of the backdoor model and the local model from a sample pair held by the malicious client as a backdoor feature and a clean feature; comparing and learning the separated back door features and the clean features, and learning the clean features for the local model by using a sample weighting strategy to train the local model to obtain a trained local model; and sending local model parameters of the trained local model to a global server, and generating model parameters of a new global model based on the local model parameters through an aggregation function. The method aims at reducing information dependence between backdoor features and clean features through comparative learning, so that local model learning is free of backdoor representation, and the robustness of a global model is improved.

The invention discloses an insurance fraud risk real-time evaluation method based on a dynamic Bayesian network. The method comprises the following steps: S1, collecting time sequence behavior data of an insurance user; s2, performing preprocessing to generate a structured observation sequence; s3, dividing into continuous time slices according to time, and constructing a time sequence sample; s4, constructing a dynamic Bayesian network; s5, calculating a posterior probability of a state variable in each time slice, and generating a sample weight; s6, executing an improved increment EM algorithm based on the sample weight, and optimizing dynamic Bayesian network parameters; s7, monitoring posterior probability entropy change of the state variable, and performing structure adjustment; and S8, the reasoning process is executed again, and a risk scoring report is generated. According to the method, the dynamic Bayesian network and the incremental EM algorithm are fused, the insurance fraud risk time sequence evaluation model is constructed, and the method has the advantages of being high in real-time performance, self-adaptive in structure, interpretable in result and the like.

The invention discloses a multi-fault diagnosis method for lithium battery sample scarcity and data imbalance, and the method comprises the steps: firstly constructing a feature extraction model based on a residual neural network, introducing an improved multi-factor imbalance index (MFI), carrying out the analysis of real-time monitoring batch feature distribution through employing a minimum spanning tree, and carrying out the real-time monitoring of the real-time monitoring batch feature distribution; therefore, the loss function and the sample weight are dynamically adjusted, and the learning stability of majority classes and the recognition precision of minority classes are both considered. On the basis, a prototype vector of a normal working condition is obtained through sample feature mean value calculation, and a prototype network (ProtoNet) is constructed to serve as an anomaly detector; after features of a test sample are extracted through the ResNet-MFII module, the Euclidean distance between the test sample and a normal prototype is calculated, if the Euclidean distance exceeds a set threshold value, it is judged that the test sample is abnormal, and detection of unknown or rare faults is achieved. The system finally outputs fault types and abnormal alarms, and high-precision recognition of multiple types of faults such as short circuit and aging of the lithium ion battery is achieved.

The invention provides a water supply network water leakage abnormity identification method based on NMF dimension reduction, and relates to the technical field of water supply network abnormity identification, and the method comprises the steps: embedding hydrophone sample data containing expert labeling information into a sample weight matrix as a row vector, and constructing a sample dimension loss item; hydrophone sample data containing industry feature priori are mapped to the feature basis matrix to serve as rows, and feature dimension loss items are constructed; combining the two loss items to construct a total loss function and solving the total loss function to obtain an optimal sample weight matrix; and performing clustering analysis based on the optimal sample weight matrix to obtain a clustering affiliation result of the hydrophone sample data so as to judge whether the water supply network to be identified has water leakage abnormity or not. According to the method, the double-constraint NMF framework fusing the expert knowledge of the sample dimension and the industry priori knowledge of the feature dimension is constructed, the accuracy and interpretability of hydrophone feature dimension reduction are considered, and the problems that an existing method is poor in interpretability, less in priori knowledge utilization and the like are solved.

The invention relates to a transferable vector quantization alignment method and device based on unsupervised domain adaptation, and the method comprises the steps: extracting the features of a source domain and a target domain of a source domain data set and a target domain data set through a feature extraction unit, and calculating an overall feature distribution loss function through searching a feature item closest to the features in a codebook; calculating a local alignment loss function through a bottleneck layer, performing classification through a classifier to obtain a source domain pseudo-label and a target domain pseudo-label, calculating a cross entropy classification loss function according to the source domain pseudo-label and a corresponding truth value label, performing normalization processing on the target domain pseudo-label, introducing mutual information to obtain a sample weight in each target domain data set, and obtaining a sample weight in each target domain data set; and calculating mutual information weighted maximization confusion matrix loss functions, and updating parameters in each module by using the loss functions until convergence to obtain a target classification architecture with cross-domain extraction, feature alignment and time sequence signal classification capabilities. By adopting the method, the identification performance of the label-free target domain can be improved.

Apparatus and method of training Machine Learning (ML) models. In an embodiment, the apparatus performs initial training of an anomaly detection model based on training samples of a training dataset over multiple epochs, where the anomaly detection model comprises a variational autoencoder (VAE). For each training sample during an epoch, the initial training comprises inputting an original data sequence of the training sample into the VAE encoder to output a multivariant distribution in latent space, sampling the multivariant distribution to generate multiple latent vectors, inputting the latent vectors into the VAE decoder to output reconstructed data sequences, and computing an estimated sample weight for the training sample. The apparatus identifies, after multiple epochs, corrupted samples from the training dataset based on the estimated sample weights, removes the corrupted samples to generate a filtered training dataset, and performs final training of the anomaly detection model based on the filtered training dataset.

A method of training a machine learning model representing association between geotechnical parameters and geophysical measurement data is disclosed. The method comprises the steps of: obtaining geotechnical training data for a first area; obtaining geophysical training data for a second area, wherein the second area is larger than and surrounds the first area; assigning sample weights to a training matrix depending on distance between the geophysical and geotechnical training data such that geophysical training data further away from the geotechnical training data has smaller weight during a learning phase compared to geophysical training data closer to the geotechnical training data; and training the machine learning model using the weighted training matrix.Unlocking insights from Geo-Data, the present invention further relates to improvements in sustainability and environmental developments: together we create a safe and liveable world.

The invention discloses a method for judging the stable state of a slope of a strip mine dump based on transfer learning. The method comprises the following steps: step 1, acquiring monitoring data of a strip mine dump slope, and dividing the monitoring data into source domain data and target domain data; 2, processing the collected monitoring data through a TCA algorithm, and mapping a processed data set to a shared feature space; step 3, in the feature space after TCA mapping, training target domain data by using a GBDT classifier, updating a sample weight at the same time, and using migration of a high-weight sample in source domain data; and step 4, utilizing the trained TCA-GBDT discrimination model to accurately discriminate the slope stability state of the target domain data, outputting a stability evaluation result, and providing a scientific basis for risk prediction and decision making of the refuse dump. The method is suitable for a small sample data set environment, the classification precision is improved, and the problem that a traditional method is low in accuracy under the condition of insufficient data is solved.

The invention discloses a multivariable time sequence anomaly detection method based on spatial-temporal feature fusion, and the method comprises the steps: outputting time sequence features through employing a series structure of a multi-scale TCN and a lightweight Transform network; constructing an adjacency matrix through a static graph learner and a dynamic graph learner, processing by using a multi-order GCN based on the adjacency matrix, and outputting spatial features; performing weighted fusion on the time sequence features and the spatial features by using a gating mechanism, and outputting a predicted value in combination with multiple residual connection; in a training process, dynamically calculating a sample weight according to a predicted residual error and calculating a weighting loss; and according to a residual error between a predicted value and a real value, generating an abnormal score through a reconstruction error calculation mechanism based on principal component analysis, and comparing a threshold value to complete abnormal detection.

The invention relates to a discrete industrial agent-based production management method and system, and relates to the field of production management, and the method comprises the steps: collecting a production prediction sample data set and a quality inspection decision sample data set, carrying out the data weight division of the two data sets, and obtaining two sample weight sets; obtaining a production prediction and quality inspection decision path array, a first prediction accuracy rate set and a decision accuracy rate set after integrated training; combining the two path arrays to obtain a discrete industrial agent array, carrying out joint optimization training, and testing to obtain a second prediction and decision accuracy set; obtaining current production basic data, inputting the current production basic data into the agent array, outputting a predicted production yield and a decision quality inspection parameter, performing compensation according to an error between the second prediction and decision accuracy set and the first prediction and decision accuracy set, and obtaining a predicted production yield and decision quality inspection parameter interval for production management. The technical problem that data interaction and business collaboration of a plurality of complex and independent scenes in the discrete manufacturing industry are difficult to realize in production management is solved.

The present application discloses a speech data processing method, device, storage medium and electronic device based on a large model, which relates to the technical field of speech data processing. The method comprises: performing a data augmentation operation on the labeled data of a labeled data set to obtain an augmented training data set; using a parsing model to classify the augmented training data set into a positive sample data set and a negative sample data set, and using the positive sample data set and the negative sample data set to generate a preference optimization data set; generating a weighted loss function according to the sample weights of the preference optimization samples and the original loss function of the parsing model, obtaining a data augmentation model that uses the weighted loss function to achieve a preset training target, and using the data augmentation model to perform data augmentation on newly input text data to obtain a target augmented data set. The above technical solution solves the technical problem of low data quality obtained by existing data augmentation methods.

The invention discloses a wind power prediction covariable offset adaptive method based on a conditional energy model. According to the method, the wind power generation power prediction model is constructed by utilizing the gated cycle unit network, and offline training of the wind power generation power prediction model is completed by adopting a sample weighting mechanism, so that the robustness of the wind power generation power prediction model to distribution change is enhanced. A conditional de-noising score matching strategy is adopted to learn the distribution difference of data in a training stage and a prediction stage through a conditional energy model, and a sample weight used for measuring the covariable offset degree is obtained based on the model. Data samples flowing in real time are stored in a replay buffer area, incremental learning is carried out on a condition energy model and a wind power generation prediction model through an online updating mechanism, and the prediction performance is kept stable. The method can effectively improve the power prediction precision and operation scheduling capability of the wind power plant under complex meteorological conditions, and has good engineering practical value and deployment flexibility.

The utility model discloses a static weighing device for a sintered ore sample, and relates to the technical field of weighing devices. The feeding device comprises a feeding machine, and a conveying belt is arranged below a discharging port of the feeding machine. In the conveying direction, a hopper is arranged below the front end of the conveying belt, the outer side of the upper end of the hopper is connected with a horizontally-arranged fixing support, a weighing sensor is installed at the bottom of the fixing support, and a supporting base is arranged at the bottom of the weighing sensor. A gate plate is arranged at the lower end of the hopper, the gate plate extends out of the hopper and is connected with a push rod for driving the gate plate to move horizontally, the push rod is connected with a push rod bracket, and the other end of the push rod bracket is mounted on the hopper. According to the utility model, the weight of a sintered ore sample can be accurately measured, and the accuracy of sample data is ensured, so that the precision and the reliability of a drum strength test are improved, and the problem of weighing deviation caused by factors such as vibration of a batching electric vibration feeder, powder residue and the like in the traditional feeding and weighing process is solved.

The present application relates to the field of electric digital data processing, and particularly relates to a joint training method and system based on real data and synthetic data, the method comprising: obtaining a mixed sample set composed of real data and synthetic data; constructing a global correlation graph based on the mixed sample set, extracting global authenticity and global closeness, and adaptively determining an optimal number of neighbors for synthetic sample quality evaluation; for each synthetic sample, constructing a neighbor subgraph by K-neighbor method using the optimal number of neighbors, extracting subgraph authenticity and subgraph closeness, and obtaining a quality score by a fusion algorithm; giving a fixed weight to real samples and a corresponding quality score to synthetic samples as a weight, constructing a weighted loss function based on sample weights, and completing joint training of the model. The present application improves the model training effect through adaptive number of neighbors, multi-dimensional quality evaluation and weighted differentiated training.

The invention discloses a tunnel engineering construction drawing design risk event safety assessment and standardized prevention and control method, which comprises the following steps: constructing a sample library matrix containing risk sources, risk events and risk prevention and control technologies at the same time, training a random forest model, and extracting risk source parameter weights; weighting the risk source parameters of the tunnel to be built by using the weight, inputting the weighted parameters into the model, and predicting the risk event probability; carrying out binaryzation by adopting a self-adaptive threshold algorithm based on F1-score to obtain a risk event prediction result; calculating multi-dimensional similarity with a sample library in combination with a Gaussian kernel function, normalizing to obtain a sample weight, and recommending an excavation method and an advanced pre-reinforcement measure according to the sample weight; and finally, a field implementation result is fed back to the sample library, and the bandwidth and the threshold are optimized by using gradient descent to realize closed-loop updating. The method reduces human intervention, improves risk assessment objectivity and prevention and control scheme standardization level, and is suitable for safety risk assessment in a complex geological tunnel construction drawing design stage.

This invention discloses a method and system for defending against big data poisoning attacks, belonging to the field of information security technology. The method includes the following steps: collecting a raw big data dataset to be labeled, and adding a credibility label column to the data to obtain a labeled big data dataset; calculating the credibility value corresponding to each sample in the labeled big data dataset, and removing samples whose credibility values ​​are less than a preset credibility threshold to obtain a high-credibility big data dataset; using the credibility values ​​corresponding to each sample in the high-credibility big data dataset as sample weights, and calculating the total loss value in combination with the loss function of the big data model to be trained; optimizing the parameters of the model to be trained based on the total loss value to obtain a big data model resistant to poisoning attacks, and using the big data model resistant to poisoning attacks to defend against big data poisoning attacks. This invention can solve the problem of training data pollution caused by data poisoning in current big data systems.

The application discloses a meter box line loss compensation method and system based on a neural network. The method first collects voltage, current, power supply end electric energy and power supply end electric energy, and calculates the abnormality degree and line loss rate of the meter box accordingly; then a data point set is constructed, DBSCAN clustering based on abnormality density weighting is performed, and multiple clustering clusters are formed; based on the abnormality distribution proportion vector similarity in each clustering cluster, the abnormality degree level is adaptively divided, and the variance of the line loss rate in each level is calculated as the line loss sensitivity; a long short-term memory neural network model is constructed, and an adaptive sample weight based on the line loss sensitivity is introduced into the loss function for training; finally, line loss compensation is performed according to the model output. The application retains key sparse samples through weighted clustering, adaptively divides abnormal levels and quantifies compensation difficulty, and combines sensitivity weighting training, so that the precision and robustness of line loss compensation in a complex scene are significantly improved.

The invention relates to a power transmission tower state determination method and device, computer equipment, a readable storage medium and a program product. The method comprises the following steps: acquiring a plurality of evaluation indexes of the structural destructive risk of the power transmission tower and a plurality of sample index values of a target evaluation index; obtaining a first weight value of the target evaluation index based on the sample index value; for any residual evaluation index except the target evaluation index in the plurality of evaluation indexes, obtaining a second weight value of the target evaluation index based on the residual evaluation index; obtaining a sample weight value of the target evaluation index and a sample state value of the power transmission tower based on the first weight value and the second weight value; training a deep learning model based on the sample index value, the sample weight value and the sample state value; and obtaining a current state value of the power transmission tower based on the trained deep learning model. According to the method provided by the invention, the accuracy of evaluating the structural destructive risk of the power transmission tower can be effectively improved.

The invention provides a construction method and system of a cancer incidence probability prediction model, electronic equipment applying the construction method and a computer readable medium storing computer program codes. The construction of the cancer incidence probability prediction model provided by the invention comprises the following steps: standardizing sample data, adding a mean value and a variance as features and clustering samples, distributing different weights for the samples according to a clustering result, then constructing a decision tree based on the weights of the samples and establishing a weighted decision forest model, and finally constructing the cancer incidence probability prediction model. Determining a final prediction result based on weighted voting classification using the total weight of each decision tree; the model realizes accurate prediction of the cancer incidence probability.

The application discloses a long video understanding-oriented adaptive key frame sampling method and system based on a video hierarchical structure, which comprises the following steps: performing hierarchical structure splitting on a long video to be processed; calling a video large language model component to generate summaries of semantic event segments, scene segments and shot segments; calling the large language model component according to a task text to calculate similarity scores of each level and comprehensive similarity scores of each shot segment; screening top K shot segments as candidate sampling units; under the constraint of a preset total sampling frame number, calculating sampling weights and frame quotas of each candidate sampling unit; and performing adaptive key frame sampling in the corresponding candidate sampling unit. The application solves the problems of poor performance of existing long video understanding methods in key frame selection tasks and consumption of a large amount of computing resources, and is suitable for video content labeling, summary generation and question and answer tasks driven by a video large language model.

The invention discloses a software deployment compatibility prediction method and system based on a LightGBM algorithm, and belongs to the technical field of data analysis. The method comprises the following steps: acquiring environmental parameters of a historical deployment task and a dichotomy deployment result; performing sample balance and weight distribution processing on the data to construct a full-amount training set; based on the training set, a two-stage LightGBM architecture is adopted to train a prediction model, a first classifier learns a global mapping relation and outputs an initial prediction probability, and after a boundary fuzzy sample set is screened out according to the initial prediction probability, a second classifier learns fine feature differences of the boundary fuzzy sample set so as to determine a final prediction result. According to the method, the problem of unbalanced category distribution is solved through cooperation of SMOTE oversampling and sample weight adjustment, the discrimination capability of boundary fuzzy samples is enhanced by using a two-stage classification mechanism, and the defect of low prediction accuracy of a traditional method in a complex heterogeneous industrial environment is effectively overcome; and the compatibility evaluation automation level and the prediction reliability before the printing industry software deployment are obviously improved.

The present invention relates to the field of structural health monitoring technology, specifically a method for predicting the settlement of a steel structure elevator shaft based on an LSTM model, comprising the following steps: analyzing triaxial inclination data, screening effective sections of inclination changes, extracting three-dimensional displacement data to identify inclination linkage mode intervals, identifying offset amplitudes to construct a sample weight distribution table, detecting jump points, setting prediction paths to obtain settlement prediction information, and comparing residuals to identify time offsets to generate structural prediction offset trend data. In the present invention, by identifying inclination inflection points and combining the differences in triaxial change amplitudes, the extraction of effective data segments is achieved, the directional accuracy of structural trend judgment is enhanced by combining the identification of the main axis displacement direction, the training sample weights are adjusted in combination with the data offset amplitude, the model's learning ability for key deformation periods is optimized, the jump density is used to set the starting point of the prediction path, the behavioral responsiveness of trend prediction is improved, and the interpretation of the prediction results is enhanced by combining the translation state residual fluctuation analysis.