71 results about "Model interpretation" patented technology

The invention discloses an engineering vehicle safety simulation and prediction system based on digital twinning, and the system comprises a data collection layer which collects multi-source heterogeneous data in real time; in the knowledge graph layer, a streaming inference engine is constructed based on an Apache Jena graph database, and an entity-relationship-attribute triple dynamic graph structure is adopted; according to the AI model layer, a physical rule serves as a loss function constraint term to be embedded into a neural network through a physical information neural network, a digital organ model concept is combined to split a vehicle into key organs for heterogeneous modeling, a simplified physical model is adopted in the core physical process, and an LSTM-AI model is adopted in external behaviors; the explanatory analysis layer is used for integrating an SHAP / LIME explanatory tool to output a visual evidence chain during fault prediction, and deploying an online incremental learning framework to allow the model to learn from new data and dynamically adjust normal range definition; and the visualization and application layer is used for performing three-dimensional visualization rendering based on WebGL or Three.js, and ensuring data transmission security through block chain evidence storage and end-to-end encryption.

The invention discloses a hepatocellular carcinoma postoperative early recurrence prediction method based on multi-modal fusion. The method comprises the following steps: firstly, integrating clinical data of a training set, a preoperative enhanced CT image and a postoperative full-view digital pathological image, and carrying out standardized correction; then, traditional image omics features and deep learning features are extracted from the CT image, cell nucleus morphological features and tumor microenvironment spatial configuration features are extracted from the pathological image, and key feature signatures are screened out through a maximum correlation minimum redundancy algorithm (mRMR) and LASSO regression in combination with clinical features. And then carrying out progressive model construction by adopting an XGBoost algorithm, sequentially establishing a clinical single-mode model, an image single-mode model, a pathological single-mode model and a multi-mode fusion model, and explaining and visualizing the models by utilizing an SHAP value and a Grad-CAM technology. Finally, the performance of the model is evaluated in a multi-dimensional mode through internal cross validation, foresight and external independent validation, risk layering is carried out based on the prediction probability, and individualized postoperative management is guided.

The invention discloses a cerebral hemorrhage patient tracheotomy risk prediction method and system based on machine learning, and the method comprises the steps: obtaining an initial clinical data set of a target patient, calculating laboratory inspection data according to a predefined rule, and constructing a composite physiological state index to generate a feature vector for prediction; inputting the feature vector for prediction into a risk prediction model pre-trained based on an ensemble learning algorithm to obtain a risk quantitative index; the model interpretation module generates an individualized prediction contribution decomposition result based on an SHAP value calculation framework, and explains the specific influence of each feature on the risk index; and finally comprehensively generating a risk prediction report. According to the method, the feature representation and model prediction capability is enhanced by constructing the composite indexes, and meanwhile, the decision process is transparent and credible by utilizing interpretability analysis, so that clinical risk assessment and decision support are effectively assisted.

The invention belongs to the technical field of water quality parameter inversion, and particularly relates to a machine learning-based suspended matter concentration remote sensing inversion method and system, and the method comprises the steps: collecting a satellite remote sensing image and synchronous actual measurement suspended matter concentration data, and obtaining a space gridding water body remote sensing reflectivity matrix through radiometric calibration, atmospheric correction and water body mask. Reconstructing and denoising through wavelet decomposition, and normalizing and standardizing the spectrum to obtain a spectrum numerical sequence; multi-scale waveband combination and differential features are constructed based on the sequence, and sensitive features are screened out by using XGBoost. A physical constraint term is constructed in combination with sensitive characteristics and a water body radiation transmission rule, and an intermediate inversion result is obtained through numerical iteration. And performing deviation compensation on an intermediate result by using a deep learning residual error, and finally performing spatial smoothing, consistency verification and high-concentration saturation optimization to obtain a high-precision and spatially continuous suspended matter concentration spatial distribution result. According to the method, efficient feature mining and physical mechanism deep fusion are realized, and the explanatory and generalization ability of the model is greatly improved.

The invention discloses a remote sensing multi-parameter integrated inversion normal form method, system and equipment based on AI-Agent. According to the method, a deep learning neural network is dynamically driven through AI-Agent, a refining mechanism (RM)-Transformer-MoE size nested model, a physical method, a statistical method and expert knowledge are coupled, a DL-C-PSK normal form is constructed, a high-precision multi-source database is established based on the normal form, an appropriate radiation transfer equation is constructed through geophysical logical reasoning, and a high-precision multi-source database is established. And inversion of parameters such as surface temperature, surface emissivity, atmospheric water vapor content and near-surface air temperature is realized. According to a causal relationship between an input wave band and an output parameter, a direct synchronous inversion or iterative inversion mode is adopted to ensure multi-parameter high-precision synchronous inversion. Wherein the core of the deep learning neural network comprises RM logic derivation, SHAP model interpretation, Transform model architecture and a Transform-MoE size nested model, so that the interpretability, the adaptability and the precision of the model are improved. Through an AI-Agent driven RM-Transform-MoE nested model, deep coupling of physics-statistics-knowledge is realized, compared with a traditional SW method, the inversion precision is greatly improved, and verification shows that the technology is suitable for the fields of global climate observation, environment monitoring and the like.

Explanatory dropout systems and methods for improving a computer implemented machine learning model are provided using on-manifold / on-distribution evaluation of dropout of key features to explain model outputs. The machine learning model is trained using a plurality of input examples, including input records with explicit dropout operators applied effectuating the removal of influence of features associated with an explanation reason class. One or more dropout operators may be stochastically applied to one or more input examples. The procedure includes on-manifold / on-distribution evaluation of the machine learning model under conditions of absence or presence of the one or more dropout operators for reliable calculation of numerical statistics associated with reason classes to yield model explanations. The training and evaluation procedures present advantages over traditional off-manifold or off-distribution perturbative explanation procedures.

The technical scheme of the invention discloses a method for judging weights of associated elements of process parameter data and quality data. The invention provides a method for analyzing relevance between process parameter data and quality data and endowing the process parameter data with weights. Aiming at the current situations that the industrial quality data acquisition sample size is small and part of values are easily identified as abnormal values when a traditional method similar to standard deviation is used, the method adopts a grouping comparison strategy when abnormal data are removed in data preprocessing, and effectively retains part of quality characteristics with large deviation. And a small amount of data still having dispute after deletion and selection is subjected to artificial expert review, so that the quality characteristics are further reserved, and the dependence on artificial deletion and selection is reduced. According to the method disclosed by the invention, in order to solve the problems that a model generated by machine learning is relatively poor in interpretability and relatively few in verification, an SHAP algorithm is adopted to carry out association construction on the model, process parameter weights are output, meanwhile, an output result is verified, and the correctness of a weight conclusion is ensured.

The invention discloses a Grad-CAM improvement method based on gradient flow direction correction and feature contribution degree redistribution, and relates to the technical field of deep learning model interpretation. Aiming at the defects that in a traditional Grad-CAM algorithm, gradient averaging causes positioning fuzziness, feature relevance is not considered in channel weight calculation, and the channel weight calculation only depends on single-level features, optimization is achieved through three key improvement steps that firstly, gradient flow direction tracking is adopted to replace global averaging to calculate channel weight, and gradient space distribution information is reserved; secondly, feature contribution degree redistribution is carried out based on channel cosine similarity, and weight superposition of redundant features is avoided; and finally, fusing the features of the intermediate layer to perform detail compensation, and enhancing the positioning precision of the thermodynamic diagram. The accuracy and robustness of model explanation are remarkably improved while the lightweight characteristic of an original algorithm is kept, and the method is particularly suitable for scenes needing high-precision explanation, such as model explanation analysis in the deep learning field.

The application is suitable for the field of model explanation, and relates to a rule mining method and device for recommending model explanation, equipment and medium. For any user-item pair in graph data, a corresponding neighborhood graph and a connected subgraph are determined, a candidate subgraph is determined from the connected subgraph, a target subgraph is determined from the candidate subgraph according to a graph evaluation score, a target graph pattern is extracted from the target subgraph, for any variable in a pattern path of the target graph pattern, a target predicate is determined from all predicates corresponding to the variable, a candidate premise condition is formed by the variable and the target predicate, a candidate explanation rule is obtained according to the target graph pattern and a candidate premise condition set corresponding to all pattern paths in the target graph pattern, and a candidate explanation rule satisfying a preset condition is determined as a target explanation rule from all candidate explanation rules. The target explanation rule reflecting the prediction principle of the recommendation model is mined from the graph data as the global explanation of the recommendation model, and the effectiveness of the explanation is improved.

The invention relates to a medical waste intelligent identification method based on improved YOLOv11, and belongs to the field of medical waste intelligent identification. According to the method, a P4-ECA module is embedded in a YOLOv11 framework in a fixed point mode, a SlideLoss dynamic weighted regression function is introduced, SCID / MCID subsets are distinguished, a Grad-CAM + + interpretable mechanism is fused, system improvement is formed in the three aspects of structural design, loss optimization and model interpretation, and high-precision, high-robustness and interpretable intelligent recognition of the medical waste is achieved.

The invention discloses a method for constructing an intelligent colorectal cancer prediction model based on mass spectrum serum proteomics, which comprises the following steps of: screening out candidate micropeptides with diagnostic potential by combining high-precision mass spectrum quantification with AI function prediction, and further constructing the model by adopting an ensemble learning algorithm. The method is rigorous in process, and by introducing an advanced normalization algorithm, integrating a feature selection strategy and targeted sample imbalance processing, the biomarker screening robustness and the prediction model accuracy are remarkably improved. Meanwhile, through model interpretive analysis, theoretical support is provided for clinical application of the marker.

The invention relates to the technical field of urban landscaping, and discloses a street tree crown coverage benefit evaluation and optimization method, which comprises the following steps: S1, collecting multi-source data of street trees in a target area; s2, performing fusion processing on the multi-source data, extracting feature parameters for evaluation, inputting the feature parameters into a pre-trained crown benefit evaluation model, and outputting a comprehensive benefit score of crown coverage by the model; wherein the crown benefit evaluation model is a machine learning model obtained by training sample data with expert score labels; and S3, based on the comprehensive benefit score and a benefit restriction factor identified through model interpretive analysis, matching and generating a targeted crown optimization scheme from a pre-constructed optimization strategy knowledge base. According to the method, the street tree crown coverage benefit evaluation and optimization of the complex urban environment can be systematically completed, the evaluation accuracy is relatively high, and the method has dynamics and operability.

The invention discloses an LIBS (laser-induced breakdown spectroscopy) multi-component quantitative analysis method based on integrated deep kernel PLS (Partial Least Squares) regression. The method comprises the following steps: S1, original spectrum kernel mapping; S2, extraction layer feature learning; S3, integrated layer feature screening; the invention provides an LIBS (Laser-induced Breakdown Spectroscopy) multi-component quantitative analysis method based on integrated depth kernel PLS (Partial Least Square) regression, and aims to develop a high-precision, transparent and light-weight depth regression model so as to solve the problems of insufficient complex nonlinear modeling capability, poor model interpretation, unbalanced multi-component prediction and the like of the existing LIBS quantitative analysis. And the end-to-end quantitative analysis of the multi-component LIBS full-spectrum data is realized.

The invention discloses a water chilling unit fault diagnosis method based on an interpretable deep learning model, and the method comprises the following steps: S1, data preprocessing and spatial feature embedding: collecting various sensor data of a water chilling unit, and carrying out the cleaning, filtering and Min-Max normalization preprocessing of the collected data; the explanation mechanism is deeply integrated into the model architecture, decision logic is presented in a visual form, the model explanation ability is improved, meanwhile, spatial feature embedding is performed according to the physical flow path of the water chilling unit, parameter spatial relevance is reserved, the understanding ability of the model to the system state is improved, high precision and practicability are achieved, and the method is suitable for popularization and application. Operation and maintenance can be assisted to quickly position faults, and the operation cost is reduced.

The invention discloses an industrial control network traffic anomaly detection method based on deep learning. The method comprises the following steps: acquiring industrial control network traffic data and performing preprocessing; performing industrial control protocol analysis to obtain an industrial control protocol semantic field, and constructing an event sequence; carrying out numerical coding on the event sequence, constructing an adjacent window pair set, and obtaining a window-level abnormal score and a time step-level difference sequence; generating an alarm trigger and constructing an abnormal interval; determining a contribution time step set, and constructing an evidence event set; and constructing an interpretation generation prompt, and obtaining a structured interpretation result by using the large model. According to the method, TimeRCD anomaly modeling based on adjacent window relative context differences and structured large model interpretation generation based on evidence events are introduced, so that unknown anomalies can be identified and field-level evidences and anomaly interval interpretation can be output under the scene of coexistence of industrial control network strong periods, station differences and process fluctuations; and the accuracy of industrial control flow anomaly detection is obviously improved.

The invention discloses an aeromagnetic airborne electronic interference feature selection method based on model interpretation force, and relates to the technical field of aeromagnetic field detection, and the method comprises the steps: firstly constructing a latent variable space through a partial least square method, so as to eliminate the multicollinearity among multi-dimensional electrical features; quantifying the actual contribution degree of each feature to the interference magnetic field based on a variable importance projection index; secondly, based on double criteria of variable importance projection and Pearson's correlation coefficients between features, through an iterative screening strategy, selecting out an optimal feature combination which is high in contribution degree and is not redundant to one another; and finally, constructing an extended Tools-Lawson compensation model based on the feature combination to carry out interference compensation. According to the method, the model is introduced to explain the real contribution of the mechanical characteristics, and the redundancy is eliminated in combination with the correlation constraint, so that the core interference source is accurately screened from the high-dimensional electrical characteristics, and the compensation effect of the airborne electronic equipment interference in the aviation magnetic survey is remarkably improved.

The invention discloses a causal-driven dynamic graph neural network interpretation generation method, which comprises the following steps of: constructing a knowledge-injected high-order structure causal model, and generating a causal soft mask of an encoding motif and hypergraph knowledge through a dynamic variational graph auto-encoder; non-causal and false components are separated and inhibited by using a multi-level mask fusion and back door adjustment mechanism; and high-order consistency loss, comparison loss and dynamic loss are introduced, so that semantic coherence, causal authenticity and time sequence evolution characteristics of interpretation are jointly ensured. According to the method, a stable and explainable high-order causal path can be identified from the dynamic graph, the fidelity of model explanation and the prediction accuracy of downstream tasks are remarkably improved, and the method is suitable for the fields of social analysis, traffic prediction, biomedical networks and the like.

Different versions of a program are executed in a first mode of operation in accordance with normal operations with only expected behavior to generate different acceptable behavior models indicative of behavior of the program. A difference between acceptable behavior models indicates different risk profiles for the different versions of the program. Risk profiles can be generated by a model explainer. An event that attempts to execute remote code or a delayed event can be detected in a second mode of operation if the event or the delayed event is not included in the acceptable behavior model. Different sequences of events in the first or second mode can be instrumented at different levels of instrumentation to generate or use the acceptable behavior model. The sequences of events can be generated from a stacktrace in the first and second modes. The instrumentation can be performed in hardware by intercepting SYSCALLs.

The invention relates to the technical field of artificial intelligence, and discloses an Internet big data analysis method and system based on artificial intelligence, and the method comprises the steps: collecting multi-source heterogeneous Internet big data, and carrying out the preprocessing of the multi-source heterogeneous Internet big data into a standardized time sequence multi-dimensional data set; potential causal variable characterization is extracted through a variational auto-encoder; constructing a dynamic causal structure map satisfying time constraint and statistical significance; estimating a quantitative causal effect by adopting dual robustness; and executing anti-fact inference and reinforcement learning driven decision optimization based on the causal atlas. The system comprises a data acquisition and preprocessing module, a causal variable extraction module, a causal atlas construction module, an effect quantification module and a decision optimization module. According to the method, intelligent decision transition from correlation analysis to causal driving is realized, the interpretability, robustness and generalization ability of the model are remarkably improved, and low-cost anti-fact deduction and automatic strategy generation are supported.

A method includes obtaining, from a matching model, for a user input, a set of candidate matches generated using a corresponding set of features. The corresponding set of features is partitioned into a first feature subset and a second feature subset. For each candidate match in the set of candidate matches, a candidate feature subset of features is individually generated. The candidate feature subset of features are selected from the first feature subset and the second feature subset for the candidate match. A multitude of candidate feature subsets corresponding to the set of candidate matches is obtained. The multitude of candidate feature subsets is clustered to obtain a first feature pattern and a second feature pattern. A large language model (LLM) generates a suggestion to modify the user input by processing the first feature pattern and the second feature pattern. The method further includes presenting the suggestion of the LLM.

Provided is a model interpretation method, an image processing method, an electronic device and a storage medium, relating to the field of artificial intelligence, in particular to the field of deep learning. The model interpretation method includes: obtaining a token vector corresponding to an image feature input to a first model; obtaining a model prediction result output by the first model; and determining, according to a combination of an attention weight and a gradient, an association relation between the token vector input to the first model and the model prediction result output by the first model, where the association relation is used to characterize interpretability of the first model.

This application provides a material vectorization method and system. The method includes: acquiring user profile data, material profile data, user behavior data, and material category structure data; constructing behavioral intent sequences and material click sequences based on user behavior data, and constructing a material graph structure based on the material category structure data; pre-training a material graph representation model by combining the material graph structure and material profile data; training a user behavior intent model by combining user profile data and behavioral intent sequences, and training a material click-through rate prediction model by combining the user behavior intent model, material click sequences, material graph representation model, and material profile data; and fine-tuning the parameters of the material graph representation model through backpropagation; after the material graph representation model gradually converges, using the fine-tuned material vectors as the vectors of the entire material system. This application offers more accurate and comprehensive vector representation, more comprehensive sequence representation, better model interpretability, and greater versatility.

The invention belongs to the technical field of traffic geographic information analysis, and discloses an urban crowd activity and space structure dynamic deduction method and system based on single traffic flow, a terminal and a storage medium. Identifying a spatial distribution mode through a flow similarity measurement method in combination with a hierarchical clustering algorithm; carrying out quantity aggregation on the identified spatial distribution mode according to grid units, and constructing a spatial grade distribution model; combining the urban built environment features with spatial grade distribution in the spatial grade distribution model, and constructing a comprehensive data set suitable for machine learning regression modeling; and based on the comprehensive data set, in combination with a machine learning model interpretation method of a game theory, quantitatively analyzing a nonlinear influence mechanism of urban built environment characteristics on different spatial distribution modes, and deducing internal relevance between urban crowd activities and spatial structures. The dynamic evolution process of the urban space structure is comprehensively realized.

This invention relates to the fields of artificial intelligence and medical clinical decision support, specifically a method and system for predicting postoperative delirium risk in elderly patients based on machine learning. The method includes: acquiring perioperative data of the patient to be predicted, including clinical indicators from the preoperative, intraoperative, and postoperative stages; preprocessing and preliminary feature screening of the data to obtain a structured feature set; constructing and optimizing a machine learning-based postoperative delirium prediction model, forming a modeling pipeline by combining multiple feature selection methods with a classifier, determining the optimal hyperparameters using random search and k-fold hierarchical cross-validation, and selecting the best pipeline based on feature stability assessment and multiple evaluation indicators; training and evaluating the performance of the final model; outputting the postoperative delirium risk prediction results and providing model interpretation. This invention is applicable to scenarios such as perioperative risk assessment of elderly patients, early warning of high-risk patients with postoperative delirium, individualized intervention plan formulation, and clinical auxiliary decision systems, providing reliable technical support for reducing the incidence of postoperative delirium and optimizing the allocation of medical resources.

The invention belongs to the technical field of SAR image classification network interpretation, and particularly relates to an interpretability analysis method and device of an SAR image classification network and a computer readable storage medium. The method comprises the following steps: firstly, calculating a Shapley value of a feature of an SAR image input into an SAR image classification network by adopting an SHAP algorithm so as to obtain an original SHAP saliency map; then, adopting a gradient-based attribution algorithm for the features of the SAR image to obtain a gradient saliency map for evaluating feature importance; and finally, fusing the original SHAP saliency map and the gradient saliency map to obtain a fused SHAP saliency map, and carrying out interpretability analysis on the SAR image classification network by using the fused SHAP saliency map. According to the method, the SHAP method is improved from the angle of increasing the gradient weight, the problem of misinterpretation encountered by a traditional SHAP and the interpretability challenge faced by the traditional SHAP in an SAR recognition task are solved, the intuition and comprehensiveness of model interpretation are further improved, and a new view angle is provided for deep analysis and optimization of an image classification model.

The application provides a blockchain-based federated learning method, device, equipment and storage medium. The method comprises the following steps: receiving training models sent by each client, calculating a plurality of model explanation values of each training model; aggregating the plurality of model explanation values of each training model into x, determining a detection score representing whether the corresponding client is a malicious client according to whether each x meets a first preset condition; determining the contribution score and the total score of the corresponding client according to whether each model explanation value meets a second preset condition, and determining the contribution proportion of each client; determining the explanation score of each client according to the detection score and the contribution proportion, and determining the trust score according to the explanation score and the reputation score; determining the training models of the top K clients in the trust score to be aggregated; the blockchain-based federated learning can be realized, the training models uploaded by the clients can be distinguished, and the quality and reliability of the federated learning aggregated model can be ensured according to the credibility of the clients.

Methods and apparatus for root cause analysis. The method includes obtaining measurement data comprising measurement data of features of a system. The method also includes generating a predicted value by applying a trained machine learning model to the measurement data and also generating a feature impact value by applying a generated machine learning model explainer to the measurement data. The method also includes updating an ontological representation of connections between features of the system and the predicted value using the generated feature impact value and outputting a proposed root cause that caused the predicted value based on the updated ontological representation.

A system may include processing circuitry and memory storing instructions, where the instructions, when executed by the processing circuitry, cause the processing circuitry to receive a first set of measurements associated with a first set of wells and generate a first well model representative of a property associated with the first set of wells. The processing circuitry may generate a well property model representative of an expected property relative to a measurement associated with a well, receive a second set of measurements associated with a second set of wells, and generate a second well model representative of a first set of predicted measurements. The processing circuitry may generate an adjusted second well model based on the well property model and the second well model, determine a second set of predicted measurements, and instruct a display to display the first set of predicted measurements and the second set of predicted measurements.