40results about How to "Guaranteed rationality" patented technology

The invention provides a one-dimensional pneumatic design method and device for an axial flow compressor, and belongs to the field of axial flow compressor pneumatic design. The method comprises the following steps: taking values of a plurality of design variables as the state of a reinforcement learning agent; the intelligent agent outputs the adjustment amount of the design variable according to the state, and the adjustment amount serves as the action executed by the intelligent agent; applying the action to the value of the current design variable to obtain an updated design variable; utilizing a one-dimensional inverse problem design program to generate a one-dimensional pneumatic design scheme of the gas compressor; utilizing a one-dimensional normal problem analysis program to output performance indexes; constructing a reward function according to the performance indexes; based on feedback provided by the reward function, a depth deterministic strategy gradient algorithm is adopted to train the intelligent agent; and performing one-dimensional pneumatic design of the gas compressor by using the trained intelligent agent. According to the method and device provided by the invention, the problems that the one-dimensional design of the axial flow compressor depends on experience, the efficiency is low, automatic optimization is difficult and the response is slow can be solved.

The invention discloses a loss diagnosis and health management system and method for a transformer, and belongs to the technical field of intelligent transformers, and the system comprises a thermal fingerprint module which is used for recording a temperature drop curve of each part of the transformer after planned power failure, and obtaining a thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves; the theoretical temperature module is used for calculating the theoretical cold body temperature of each part of the transformer based on the real-time environment temperature and the thermal inertia fingerprint when the transformer runs; the net temperature rise module is used for acquiring the actual temperature value of each part of the transformer and subtracting the corresponding theoretical cold body temperature from the actual temperature value; weak thermal characteristic signals caused by insulation aging and loss abnormity of the transformer can be accurately extracted from strong background noise such as load fluctuation and environment temperature through technologies such as thermal inertia fingerprint construction and net temperature rise separation, early diagnosis of progressive health deterioration is realized, support is provided for intervening potential faults in advance, and the fault diagnosis accuracy is improved. And the passive situation of post-treatment of operation and maintenance is changed.

The application provides a control method and a control device of compressor operation frequency and a water heater. The control method comprises the following steps: calculating the suction pressure attenuation ratio of the compressor; judging the frost degree of the heat exchange equipment according to the suction pressure attenuation ratio of the compressor; and dynamically setting the initial operation frequency of the compressor according to the frost degree of the heat exchange equipment. The control method further comprises the following steps: comparing the initial operation frequency of the compressor with the historical operation frequency of the compressor recorded in a database; and controlling the actual operation frequency of the compressor according to the comparison result. The control method can guarantee the rationality of the actual operation frequency of the compressor, solve the problem of low defrosting efficiency caused by the fixed operation frequency of the compressor during the defrosting process, and improve the energy efficiency.

The application discloses a kind of collaborative robot job parameter dynamic optimization method, comprising: real-time acquisition of the multimodal sensor data of collaborative robot in the process of job, multiple job parameters are analyzed based on job quality evaluation model to identify key parameter set from it;Multi-objective function is constructed with job time, end force as optimization target, and multi-objective function is converted into single-target optimization function by weighting sum method;Key parameter set is input to machine learning model, and single-target optimization function is used as optimization target to carry out iterative optimization, in each iteration, the standard deviation σ of the current learning result of machine learning model is obtained, according to the comparison result of standard deviation σ and preset critical value s, dynamically select optimization strategy, when σ≥s, new parameter set is generated using random exploration strategy, when σ<s, new parameter set is generated using guided optimization strategy of machine learning model, and the optimized parameter set is output.

The application provides a method and system for multi-party secure calculation of coefficient of variation, belonging to the field of multi-party secure calculation, and used for providing a new scheme for achieving the task of securely calculating the coefficient of variation of a data group, wherein the multi-party includes an initiator of a calculation task, a referee and a plurality of participants, the participants and the initiator hold private data, the initiator and the referee are two different parties, and the method comprises the following steps: the participants pre-process the data, the referee generates a public key and a private key based on a self-developed homomorphic encryption algorithm, and sends the public key to the participants, so that the private data of all the participants is encrypted, all the participants collect the encrypted data to the initiator, the initiator calculates an encrypted sum, the referee decrypts the sum and sends it to the initiator to obtain a mean value, all the participants calculate variance elements according to the mean value, and the variance elements are segmented, recombined and collected to the initiator, so that the initiator can calculate a standard deviation and then calculate a coefficient of variation.

This invention relates to the field of medical information technology, and in particular to a method and system for generating clinical multi-agent task sequences. The method includes the following steps: acquiring and structuring core clinical objectives, which include patient basic information, clinical indicators, and disease management goals; decomposing the core clinical objectives into several structured sub-objectives with dependencies and execution order based on a clinical pathway template library and a medical knowledge graph; for each structured sub-objective, splitting it into multiple minimum executable clinical sub-tasks according to a clinical sub-task decomposition rule library, and defining the execution subject type, input / output, and dependencies for each clinical sub-task to construct a sub-task-level directed acyclic graph; integrating the sub-task-level directed acyclic graphs, identifying dependencies across structured sub-objectives, generating a global clinical task directed acyclic graph, and performing dependency verification and clinical logic consistency verification.

The invention relates to a self-adaptive management method and system for a database plan cache, and the method comprises the steps: firstly judging whether the plan cache is hit or not after a database receives an SQL statement, and triggering a hard analysis generation plan and caching the plan when the plan cache is not hit; if the cache is hit, further judging whether a corresponding multi-execution plan exists or not; and when only a single execution plan exists, analyzing the relationship between the number distribution of execution tuples and the number of execution times by adopting an execution efficiency algorithm to dynamically judge whether the data is inclined. And when a plurality of execution plans exist, entering a multi-plan cache management stage, selecting the execution plans according to a matching result of the current binding variable predicate selection rate and the cache plan historical selection rate, and generating and caching a new plan when the matching fails. According to the method, the data skew is accurately identified by executing the efficiency algorithm, intelligent matching and multiplexing of plans are realized by means of a multi-plan cache management mechanism, and the overall performance of the database in a data skew scene is improved.

The invention discloses an analog signal range judgment and switching system with calibration and self-diagnosis functions, which relates to the technical field of analog signal monitoring and comprises an analog signal acquisition module, an analog signal range analysis module, an analog signal range switching module and a database, the method comprises the following steps: firstly, analyzing the influence of different environment temperature intervals on the optimal linear interval of the ADC to obtain the optimal linear interval of the ADC in the different environment temperature intervals, determining the optimal linear interval of the ADC according to the current environment temperature, and judging whether an analog signal falls into the optimal linear interval of the ADC and whether the current range gear loses efficacy; and then the target range gear is obtained by analyzing the influence among the range gears under different gear working conditions, and the current gear is switched to the target range gear, so that the effectiveness of range analysis and the precision of ADC sampling are guaranteed, and the rationality of range switching is also guaranteed.

The application provides a multi-round reasoning network traffic prediction method and device based on size model cooperation, which comprises the following steps: inputting historical traffic data into at least two time series prediction models, obtaining prediction results and calculating comprehensive confidence based on prediction errors, and generating a prediction effect ranking; inputting historical data, prediction results and confidence into a large language model, and evaluating and ranking the prediction effect of each model by the large language model; comparing the two rankings, and if they are consistent, calibrating and optimizing the large language model, and if they are inconsistent, adopting an independent adjustment strategy to enhance the sensitivity of the prediction results of the model with the best ranking, and obtaining target prediction results. Through the two-way comparison of the independent ranking of multiple models and the evaluation ranking of the large language model, the large language model can be isolated when the rankings are inconsistent, and an independent adjustment strategy is adopted, thereby effectively avoiding the direct impact of the reasoning deviation of the large language model on the prediction results, and improving the robustness and accuracy of network traffic prediction.

The application discloses a kind of based on multi-sensor's grid map two-way asynchronous construction method and system, the method includes: obtaining three-dimensional point cloud data, RGB image and depth map;Based on the space corresponding relation of RGB image and three-dimensional point cloud data, three-dimensional point cloud data is semantically labeled;Three-dimensional point cloud data, RGB image and depth map are timestamped alignment, coordinate transformation and filtering processing, obtain final point cloud data, with semantic point cloud data point cloud data is spatiotemporal alignment generation fusion point cloud data as semantic grid map input;Construct semantic grid map, adopt the two division strategies of front view map and rear view map, by independent thread asynchronous update;Establish motion object database and predict dynamic object motion trajectory based on time sequence prediction algorithm, update in semantic grid map synchronously.The application effectively guarantees the intelligent device obstacle avoidance success rate and path planning rationality, guarantee the operation safety and efficiency under complex environment.

The embodiment of the invention provides an insurance premium auditing method and device, equipment, a storage medium and a program product. Belongs to the technical field of data processing. The method comprises the following steps: receiving a to-be-converted rate table file sent by a configuration terminal and insurance type information of a target insurance type; converting the to-be-converted rate table file into a to-be-used rate table; determining a compliance premium interval and an authorized premium interval according to a to-be-used rate table, a target insurance type identifier, a target insurance amount, a target risk factor and target limit claim-free information in the insurance type information; collecting real-time reference data corresponding to the target risk factor; determining a to-be-used authorized premium interval according to the real-time reference data and the authorized premium interval; determining an audit result corresponding to the target insurance premium according to the to-be-used authorized insurance premium interval and the compliance insurance premium interval; and outputting an auditing result. The method solves the problems that the efficiency of manual insurance premium auditing is low, and the insurance premium auditing result is difficult to reflect the real-time risk change.

The invention provides a distribution network non-power-cut operation digital decision and scheme generation method and system based on a live-action three-dimensional model. The method comprises the following steps: firstly, constructing a three-dimensional digital twinborn scene; receiving user job task information, positioning to a position corresponding to the scene, and calling a job method knowledge base to screen a preliminary candidate job method; and then simulating candidate method implementation in the scene, screening out qualified candidate methods through dynamic safety distance verification and feasibility analysis, generating a recommendation scheme sequence through quantitative evaluation comparison and selection, generating a preliminary job implementation scheme report according to template filling information, and finally receiving user auditing and modifying confirmation information and storing a final scheme. According to the method, the three-dimensional space relation of the line corridor is accurately reproduced by means of the three-dimensional digital twinning scene, quantitative data support is provided for operation method selection through multi-link simulation, verification and evaluation, information blind areas and misjudgment risks are eliminated, decision subjectivity is reduced, and rationality and safety of an operation scheme are guaranteed.

The application discloses a method for enhancing position monitoring of a closed space PNT network and a high-precision map, and relates to the field of wireless positioning. The method comprises the following steps: establishing a semantic unit grid of a target closed space; establishing a mapping relationship database of semantic information of any one entity object element and position information of the semantic unit where the entity object element is located; acquiring a grid code of a semantic unit where a target entity object is located; and obtaining position semantic information of the target entity object based on the grid code of the semantic unit and the mapping relationship database. According to the technical scheme, the grid code of the semantic unit where the user is located is calculated only once, the semantic information is obtained by reading the mapping relationship database index based on the grid code of the semantic unit, and the calculation amount and response waiting time of the user position monitoring information are greatly reduced. Furthermore, the rationality of the user position monitoring information is improved, and the calculation amount of the generated user position density monitoring information is reduced.

The application discloses a kind of machine learning training topology generator and the method for generating topological discriminator model, comprising the following steps: with matrix pair form characterizing real topological data, and pre-processing;Topology generator is constructed, and topological pseudo-matrix pair data is generated;Topological discriminator is constructed to identify the source of topological data;The topological generator and the topological discriminator are based on the adversarial training of generative adversarial network;Low-voltage area topological matrix pair data is generated, and is converted into visual topological graph.The application realizes the coevolution of topological generator and topological discriminator by constructing topological generator and topological discriminator and through the adversarial training of both, and then enables topological generator to automatically generate a large number of topological data in accordance with real station area statistical distribution and engineering rationality from simple input, with the technical effects of high generation quality, high automation efficiency, guaranteeing physical rationality, solving data scarcity problem and strong flexibility and adaptability.

The application discloses an indoor situation generation method and device combining thin-plate spline interpolation and natural neighborhood interpolation and a medium, and relates to the field of electric digital data processing.The method comprises the following steps: acquiring original sampling data; performing grid division, merging and extracting data falling into the same grid to obtain a first sample data set; generating preliminary interpolation data by using thin-plate spline interpolation; extracting supplementary sampling point data from the preliminary interpolation data in an area not covered by the original sampling data to obtain a supplementary sample data set; performing grid division and merging to obtain a second sample data set; combining the supplementary sample data set and the second sample data set, filtering the supplementary data, and obtaining a target sample data set; and generating final situation data by using natural neighborhood interpolation.The application can effectively improve the situation generation quality under the condition of uneven sample distribution by filling in sparse area samples through thin-plate spline interpolation and finely processing the samples through natural neighborhood interpolation.

The invention discloses an acoustic radar near-surface wind speed inversion method and system based on entropy weight feature fusion, and the method specifically comprises the steps: carrying out the preprocessing of acoustic radar frequency spectrum data, and obtaining the preprocessed frequency spectrum data; based on the preprocessed spectrum data, constructing a Doppler window dynamic migration model based on turbulence dynamics power law distribution; on the basis of a Doppler window dynamic migration model, candidate spectrum peaks are selected at each height layer, a feature space is constructed, feature weights are dynamically distributed by using an information entropy theory, and an optimal spectrum peak is selected; introducing a symmetry constraint optimization strategy to the optimal spectrum peak, detecting the profile symmetry of the wind spectrum in the symmetric direction, and calibrating an asymmetric result to obtain a target wind speed inversion result; and comparing and evaluating the target wind speed inversion result and the wind speed of the anemometer tower as a reference to complete inversion result verification. According to the method, the wind speed inversion precision of the acoustic radar in a complex environment is effectively improved, and better evaluation and efficient utilization of wind energy resources are promoted.

The utility model relates to punching equipment technical field, and disclose a new energy automobile frame processing is with punching equipment, including numerical control laser cutting machine, the workbench of numerical control laser cutting machine places the locating disc, the lower extreme of locating disc is close to both ends place and rotates and installs the support shaft, the left and right both ends of support shaft all are fixedly installed with double conical roller, the front and back both ends of numerical control laser cutting machine all are fixedly installed with the baffle, the inside of slide groove all slide and install square bearing, the inside between square bearing all are fixedly installed with two -way screw rod, the mobile block between in the similar one end all are fixedly connected with the angle and resist piece. The device can be more labor -saving and fast to move to the workbench with the cooperation of double conical roller and track, and meanwhile can fix the same batch chassis frame on another set of locating disc in advance, realizes the feeding and discharging fast, avoids lifting to the workbench and then carries out the tedious tooling positioning operation, improves the efficiency of overall processing procedure.

The invention relates to the technical field of ecological restoration, in particular to an ecological restoration monitoring method and system for a coal mining subsidence area. Obtaining soil recovery state time sequence information; when the soil recovery state at the end point moment of the time sequence information of the soil recovery state meets a soil target state, the time sequence information of the soil recovery state and vegetation planting parameters are input into a vegetation recovery state prediction network, and the time sequence information of the vegetation recovery state is obtained, the vegetation recovery state prediction network is a black box model constructed based on a long-short term memory neural network; and when the vegetation recovery state at the end point moment of the vegetation recovery state time sequence information meets a vegetation target state, performing ecological recovery monitoring on the soil modification coal mining subsidence area executing the soil modification control parameters by taking the soil recovery state time sequence information and the vegetation recovery state time sequence information as monitoring baselines. According to the method, scientificity and effectiveness of cultivated land protection and ecological restoration of the coal mining subsidence area can be improved.

The invention provides a behavior prediction model construction method and a behavior prediction method in a manual control stage of an automatic driving vehicle, and relates to the technical field of intelligent automobile and man-machine cooperative driving, and the method comprises the steps: converting multi-dimensional structured data into a sequence text which can be analyzed by a large language model based on a semantic template, and inputting the sequence text; taking the converted semantic sequence as input, constructing a multi-modal large model guided generative manual control behavior prediction model, and outputting a manual control behavior sequence; and taking a multi-modal large model guided generative manual control behavior prediction model as a teacher model, and performing knowledge distillation on the teacher model to obtain a manual control stage behavior prediction model based on cross-scene knowledge distillation migration. According to the method, the manual control behavior prediction model guided by the multi-modal large model is constructed, and the construction of the manual control behavior prediction model adaptive to different driving scenes is completed by means of a cross-scene knowledge distillation migration technology.

This application provides a resource sharing method, resource sharing device, and storage medium, relating to the field of communication technology, for improving the utilization rate of network resources in access network equipment. Applied to a first blockchain node, which is an access network device contracted with the blockchain system, the method includes: generating a first block based on the network resource information of the first blockchain node, and publishing the first block to the blockchain system; predicting the amount of shareable network resources of the first block within a target time period based on the historical network resource information of the first block; the target time period being the period after the current time; and sending a sharing request to the blockchain system when the amount of shareable network resources exceeds a preset threshold; the sharing request information indicating the amount of shareable network resources of the first block within the target time period.

The invention discloses a fine-grained multi-modal knowledge enhanced image generation method and device, and the method comprises the steps: obtaining the text features of an optimized and expanded cue word through a pre-trained text encoder, obtaining the visual features of a corresponding subject image through an image encoder, and carrying out the recognition of the subject image according to the obtained text features and visual features. Splicing is carried out in the feature dimension direction; using a trainable MLP to carry out dimension reduction on the splicing features, converting the features after dimension reduction into embedding conditions, and inputting a reference image into a reference branch to optimize attention parameters; performing multi-dimensional quality evaluation on the generated image, comprehensively generating relevancy between the image and a text cue word and semantic consistency between the image and a domain knowledge graph to output a final quality evaluation score, comparing the final quality evaluation score with a preset threshold value, and judging whether to trigger an introspection agent mechanism by referring to the current feedback frequency; and sending the image hidden space representation obtained from the output of the trunk DiT branch into an image decoder, and restoring to obtain a final generated image. The device comprises a processor and a memory.

The application relates to a distributed photovoltaic output control method, device, equipment, medium and program product. The method comprises the following steps: collecting the running state data of a current time window of a transformer area and a distributed photovoltaic access point; calculating a transformer area safety budget value in a next time window; performing consistency constraint processing on the transformer area safety budget value to obtain a final safety budget value; obtaining a final predicted safety margin of a preset future time range according to a historical running state data sequence; determining a safety control parameter according to the relationship between the final predicted safety margin and a preset safety margin threshold; and generating a control instruction according to a preset safety budget allocation rule, the final safety budget value, the safety control parameter and the final predicted safety margin, and instructing a transformer area side terminal to control the output of a distributed photovoltaic unit in the transformer area based on the control instruction. The method can control the output of the distributed photovoltaic unit in a timely manner.

The application discloses a water quality anomaly identification method and system for a water supply plant, relates to the technical field of water quality anomaly identification, and comprises the following steps: constructing a VARX model containing endogenous variable lag terms and exogenous variable lag terms, and training the VARX model by taking the minimization of a convex loss function as a target, wherein the convex loss function comprises a sample weight balancing term for balancing the fitting priority of normal and abnormal samples and a process output constraint term conforming to the process practice; analyzing historical sample monitoring data based on a box-and-whisker plot to determine the normal fluctuation interval of various variables; inputting real-time monitoring data into the trained VARX model to obtain water quality index predicted values; and comparing the water quality index predicted values with the normal fluctuation interval to determine whether the water quality is abnormal. The application guarantees that model parameter estimation converges to a unique global optimal solution under the premise of improving anomaly identification precision, and improves the stability of the model.

The present disclosure provides a knowledge mastery evaluation method, a question recommendation method, a device, equipment and a medium, wherein the knowledge mastery evaluation method comprises: obtaining a question difficulty and discrimination of a question of a target knowledge point, and a correct or incorrect result of an answer of a user to the question of the target knowledge point; obtaining a plurality of preset simulation capability values and a preset weight corresponding to each simulation capability value; obtaining a maximum likelihood estimation value corresponding to each simulation capability value under the target knowledge point according to the question difficulty, the discrimination and the correct or incorrect result of the answer; performing weighted average processing on the plurality of simulation capability values based on the preset weight corresponding to each simulation capability value and the maximum likelihood estimation value corresponding to each simulation capability value under the target knowledge point, to obtain a capability average of the user; and the capability average is used to represent the mastery degree of the user to the target knowledge point. The present disclosure can effectively guarantee the accuracy and reliability of the mastery degree evaluation result, and help to further guarantee the rationality of the question recommendation on this basis.

The invention belongs to the technical field of water meter assembly. The invention provides a single-workbench water meter assembly method and system based on body intelligence. A four-layer framework of perception, prediction, planning and execution is adopted. The sensing module outputs a water meter part state and a structured scene map based on the 3D scene reconstruction network; the prediction module generates collision risk and assembly execution time prediction based on the Transform multi-task sequential network; the planning module generates an optimal assembly action sequence through hierarchical reinforcement learning; and the execution layer executes operation in a position-force hybrid control mode, and feeds back data in real time to form a closed loop. According to the scheme, the problems that traditional manual assembly is low in efficiency and existing automatic equipment lacks self-adaptive capacity are solved, the defects that sudden working conditions are difficult to deal with and multi-model adaptation is poor are overcome, the assembly efficiency and consistency are improved, manual intervention is reduced, assembly interruption and part damage are avoided, and flexible automation of the whole water meter assembly process is achieved.

The application discloses a virtual role behavior action prediction method and system based on a neural network, which introduces physical rules in the form of executable codes and semantic constraints in the form of hot update rule tables for double checking after model output; under the premise of not retraining the model, the mechanism of combining the high generalization ability of the neural network with the certainty of the rule engine effectively guarantees the legality, rationality and authenticity of the virtual role behavior; through integrated Grad-CAM or SHAP feature attribution analysis, a logical link from environment perception to behavior execution is established, so that the developer can accurately quantify the contribution of position and posture factors to the decision, and through the introduction of controlled disturbance at the input end, the stability is screened, so that the parameters can still remain stable under abnormal interference, and the role jitter and collapse caused by network fluctuation or abnormal operation online are avoided, and the engineering reliability of the system in a complex interactive environment is greatly enhanced.

This application discloses an adaptive optimization method and system for thermal coupling parameters of automotive backlight panels, relating to the field of parameter simulation optimization technology. The specific implementation scheme is as follows: A full-envelope operating condition script is compiled; a thermal model and thermal coupling parameter set are imported; thermal simulations are automatically run in batches sequentially using the full-envelope operating condition script; raw datasets for each operating condition segment are collected; risky operating condition segments are screened; thermal coupling parameter perturbation groups are applied to each risky operating condition segment; high-risk operating condition segments are identified; parameter inversion is performed on each high-risk operating condition segment; and the thermal coupling parameter update strategy for each high-risk operating condition segment is determined, along with the optimal thermal coupling parameter version. This invention can improve the prediction accuracy of the thermal model under complex or extreme operating conditions, thereby avoiding problems such as underestimated junction temperature peaks, distorted hotspot locations, and drifting temperature rise rates. This provides a scientific and reliable data basis for the structural design, safety margin setting, and temperature control strategy optimization of automotive backlight panel radiators.

This application discloses an interactive control method and system based on virtual reality, relating to the field of monitoring and analysis technology. It includes a multimodal motion capture module for capturing limb spatial coordinate data, joint rotation angle data, and muscle activation signals corresponding to the target user; a motion intent parsing module, including a temporal convolutional neural network model, comprising an input layer and an output layer. The input layer receives skeletal joint point sequences, angular velocity data, and electromyographic feature vectors from the multimodal motion capture module, while the output layer extracts spatiotemporal features through a three-layer dilated causal convolution and connects to a classifier to generate initial motion data corresponding to the target user; a correction module constructs a human joint angle constraint library, judges the initial motion data corresponding to the target user, and outputs correction commands; and an environment adaptation module adjusts the environment in the virtual scene. This application improves interaction efficiency.

The invention relates to the technical field of clothing optimization management, and discloses an optimization management method and system based on clothing positioning price tag displacement monitoring, and the method comprises the steps: building a base station map, the base station map records position information of a plurality of electronic price tags and corresponding clothing information, a plurality of detection areas are established according to the base station map, the electronic price tags periodically report position data to a server according to the detection areas where the electronic price tags are located, and the server sends the position data to the server based on the number of movement times of the electronic price tags in a preset time period. And calculating a selected frequency corresponding to each electronic price tag, and sorting all the price tags according to the selected frequencies. According to the method, whether a user is interested in clothes or not is calculated and inferred by establishing a selected frequency algorithm, all price tags are ranked from high to low according to the values, and follow-up actions are triggered if the scores are high and the price tags are judged as high-popularity commodities, so that display optimization and collocation recommendation are assisted to be realized.

The application belongs to the technical field of data processing and customer identification, and more particularly relates to a milk powder new customer identification method based on a time window and storage fusion. Through standardized processing of multi-source data collection and business time extraction, system record time interference is excluded, and the time sequence dislocation problem of multi-source heterogeneous data is solved from the source. A unified consumer behavior timeline is constructed through event causal relationship inference, further calibrating the time sequence of behavior events and ensuring the authenticity of customer consumption tracks. Through equivalent modeling of storage and extraction behaviors, the purchase and consumption logic under the special consumption mode of the milk powder industry is accurately restored. Finally, through dynamic time window backtracking and weighted judgment, accurate identification of new customers is realized. The accuracy of milk powder new customer identification is significantly improved, avoiding misjudgment and missed judgment, ensuring the rationality of new customer identification results, and helping enterprises optimize customer hierarchical management, improve marketing resource utilization rate and customer satisfaction.