An electronic pricing scale big data acquisition processing and intelligent analysis method based on NFC and AI

Abstract

The present application relates to big data processing and intelligent analysis technical field, specifically to a kind of electronic pricing scale big data acquisition processing and intelligent analysis method based on NFC and AI;Its core is: through pre-training evaluation model to original transaction data stream is inferred, according to the preset value and risk threshold value division processing level;Based on the level, in combination with topological strategy library and system load, according to utility and cost weight selection optimal processing topology and execute;Loss function is constructed by collecting processing delay, resource consumption and analysis efficiency, and threshold value and weight parameter are reversely dynamically updated;The present application realizes the change from rigid processing to dynamic stratification and adaptive topology optimization based on data instantaneous value and risk.

Description

Technical Field

[0001] This invention relates to the field of big data processing and intelligent analysis technology, specifically to a method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI. Background Technology

[0002] In big data applications of electronic price-computing scales based on NFC and AI, the system needs to collect and process massive amounts of raw transaction data streams in real time. Existing technical solutions generally adopt fixed data processing procedures. This rigid processing mode cannot dynamically respond according to the instantaneous value and analytical risks of the data stream, nor can it allocate resources based on the current real-time load of the system. This mode leads to a serious disconnect between computing resource allocation and data value: on the one hand, valuable computing resources may be consumed in large quantities on low-value routine data, resulting in resource waste; on the other hand, critical data streams with high value or high risk may not be able to obtain sufficient computing resources due to resource competition, resulting in analysis delays or insufficient analysis depth, making it difficult to guarantee the real-time nature of critical analysis.

[0003] Therefore, how to construct an adaptive intelligent analysis method that enables the system to dynamically evaluate the value of data and intelligently allocate processing resources in combination with real-time system load to achieve the best balance between analytical efficiency and resource consumption is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI. Specifically, the technical solution of this invention includes:

[0005] S1. Collect the raw transaction data stream of the NFC electronic price-computing scale in real time, extract key features, and construct the raw feature vector; use a pre-trained evaluation model to infer the raw feature vector and output the evaluation vector; preset value threshold and risk threshold, compare the evaluation vector with the value threshold and risk threshold to obtain the processing level;

[0006] S2. Based on the processing level and combined with the preset topology strategy library and real-time system load, evaluate the expected utility and expected cost of each topology in the topology strategy library; select the optimal processing topology according to the utility weight and cost weight; instantiate the data processing link according to the optimal processing topology, perform data processing of the original feature vector, and output the analysis results; monitor and record the topology reconstruction delay and actual resource consumption, as well as the analysis results.

[0007] S3. Collect processing level and topology reconstruction delay, and collect actual resource consumption and analysis results; calculate ineffective analysis resource consumption rate and analysis efficiency; combine ineffective analysis resource consumption rate, topology reconstruction delay, and analysis efficiency to construct a loss function; use the loss function to update the target parameter vector, which includes value threshold and risk threshold, as well as utility weight and cost weight, to dynamically optimize the determination of processing level and the selection of optimal processing topology.

[0008] Preferably, S1 includes:

[0009] S11. Use a pre-trained evaluation model to reason about the original feature vector and output an evaluation vector, which includes an instantaneous value component and an analytical risk component.

[0010] S12. Determine whether the risk component is greater than the risk threshold. If so, the treatment level is high risk.

[0011] S13. If the risk component is not greater than the risk threshold, then determine whether the instantaneous value component is greater than the value threshold. If so, the processing level is high value level.

[0012] S14. If the risk component analyzed is not greater than the risk threshold and the instantaneous value component is not greater than the value threshold, then the treatment level is the normal level.

[0013] Preferably, S2 includes:

[0014] S21. Extract each topology from the topology strategy library. The topology defines the combination of data preprocessing operators, the AI ​​model inference path, and the data aggregation granularity.

[0015] S22. For each processing level, evaluate the expected utility of each topology;

[0016] S23. Evaluate the expected cost of each topology for real-time system load;

[0017] S24. Combining utility weight, cost weight, expected utility, and expected cost, solve for utility maximization to select the optimal processing topology.

[0018] Preferably, S2 further includes:

[0019] S25. Receive the optimal processing topology instruction, and the resource scheduler allocates computing resources, loads the required operators and AI models, and constructs the instantiated processing link;

[0020] S26. A priority queue mechanism based on processing level is adopted to route the original feature vector to the instantiated processing link;

[0021] S27. Perform the processing and output the analysis results.

[0022] Preferably, the monitoring and recording steps include:

[0023] S28. Monitor used resources and total system resources, calculate and generate system load;

[0024] S29. Record the output time of the optimal processing topology instruction, and the time when the instantiated processing link is completed and running stably;

[0025] S30. Calculate the time difference between the output time and the stable operation time to generate the topology reconfiguration delay.

[0026] Preferably, S3 includes:

[0027] S31. Identify and determine the data stream set as a normal level;

[0028] S32. Obtain the actual resource consumption of the data stream at the normal level, and the minimum baseline resource cost for processing the data stream at the normal level;

[0029] S33. Calculate the difference between the actual resource consumption and the minimum baseline resource cost, and compare it with the total resource consumption of the system processing all data streams to calculate the invalid analysis resource consumption rate.

[0030] S34. By comparing the analysis results and processing levels, assess the achievement of the analysis objectives to calculate the analysis effectiveness.

[0031] Preferably, S3 also includes:

[0032] S35, Preset waste weight, delay weight, and performance weight;

[0033] S36. Based on waste weight, delay weight, and efficiency weight, and combined with ineffective analysis resource consumption rate, topology reconstruction delay, and analysis efficiency, construct a loss function;

[0034] S37. Calculate the gradient of the loss function with respect to the target parameter vector;

[0035] S38. Update the target parameter vector using gradients based on the preset learning rate.

[0036] Preferably, the original transaction data stream includes product code, weight, unit price, timestamp, and merchant ID, and the original feature vector is constructed based on the extraction of the original transaction data stream.

[0037] Preferably, after the target parameter vector is updated, the updated value threshold and risk threshold are used for subsequent determination of the processing level in S1; the updated utility weight and cost weight are used for subsequent selection of the optimal processing topology in S2.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. This method uses a pre-trained evaluation model to evaluate the instantaneous value and analytical risk of the original transaction data stream in real time and classifies it into processing levels. This overcomes the rigidity of the data processing process in the existing technology and realizes the fine-grained stratification of data with different values ​​and risks.

[0040] 2. This method can dynamically evaluate the expected utility and cost of each processing topology based on the processing level of the data stream and the real-time system load status, combined with a preset topology strategy library, thereby selecting the optimal processing link; this solves the problem of the disconnect between computing resource allocation and data value, and ensures that high-value or high-risk data streams obtain sufficient computing resources in resource competition;

[0041] 3. This method, by matching the optimal processing topology to data streams of different levels and adopting a priority queue mechanism based on processing level, can concentrate valuable computing resources on processing high-value and high-risk critical data, while taking low-cost processing measures for regular data streams, avoiding resource waste, significantly shortening the latency of critical analysis, and ensuring real-time performance.

[0042] 4. This method constructs a closed-loop adaptive optimization mechanism; by monitoring topology reconstruction delay, actual resource consumption and analytical performance, and constructing a loss function, the system can automatically and dynamically optimize core decision parameters such as value threshold, risk threshold, and utility and cost weights, achieving the best balance between analytical performance and resource consumption. Attached Figure Description

[0043] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0044] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0046] Example 1:

[0047] Please see Figure 1 A method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI, the specific steps of which include:

[0048] S1. Collect the raw transaction data stream of the NFC electronic price-computing scale in real time, extract key features, and construct the raw feature vector; use a pre-trained evaluation model to infer the raw feature vector and output the evaluation vector; preset value threshold and risk threshold, compare the evaluation vector with the value threshold and risk threshold to obtain the processing level;

[0049] S2. Based on the processing level and combined with the preset topology strategy library and real-time system load, evaluate the expected utility and expected cost of each topology in the topology strategy library; select the optimal processing topology according to the utility weight and cost weight; instantiate the data processing link according to the optimal processing topology, perform data processing of the original feature vector, and output the analysis results; monitor and record the topology reconstruction delay and actual resource consumption, as well as the analysis results.

[0050] S3. Collect processing level and topology reconstruction delay, and collect actual resource consumption and analysis results; calculate ineffective analysis resource consumption rate and analysis efficiency; combine ineffective analysis resource consumption rate, topology reconstruction delay, and analysis efficiency to construct a loss function; use the loss function to update the target parameter vector, which includes value threshold and risk threshold, as well as utility weight and cost weight, to dynamically optimize the determination of processing level and the selection of optimal processing topology.

[0051] This invention discloses a method for big data acquisition, processing and intelligent analysis of electronic price-computing scales based on NFC and AI. Its purpose is to solve the problems of rigid data processing flow and inability to dynamically allocate resources according to the instantaneous value of data streams and system load in the existing technology.

[0052] The system collects raw transaction data streams from NFC electronic price-computing scales in real time; after collection, the system extracts key features from the data streams and constructs raw feature vectors. This process is the foundation for achieving AI analysis.

[0053] Using a pre-trained evaluation model For the original feature vector Perform inference; evaluate the pre-trained model. The purpose is to conduct a rapid online assessment of the potential value and analytical risks of data streams; in this embodiment, It can be a lightweight neural network with its model parameters It is obtained through offline training on massive amounts of historical transaction data; to obtain supervised labels for offline training, those skilled in the art can use the following methods to label the historical data:

[0054] For instantaneous value components Post-hoc analysis is used to assess the value of historical transaction data and normalize it. Value; for analyzing risk components By associating known historical risk events, these events are marked as high-risk and normalized to... value;

[0055] Inference output evaluation vector ;Should It is a two-dimensional vector whose components include instantaneous value components. and analysis of risk components Both were normalized.

[0056] The system is based on a preset value threshold. With risk threshold , will evaluate vector Compare with these two thresholds; value threshold and risk threshold These are the boundaries for the system to achieve dynamic classification. They are not fixed, but are determined by the target parameter vector in the subsequent step S3. Dynamically provided and updated;

[0057] The processing level is obtained after comparison. This processing level It is a discrete signal, such as high risk. High value ,conventional This is used to guide S2's decision-making;

[0058] The system is based on the processing level input from S1. And combined with the preset topology strategy library With real-time system load Dynamic decision-making; Topology strategy library Multiple processing topologies are predefined in the code. Each topology Corresponding to a data processing flow; system load It is a real-time changing parameter that reflects the current workload of computing resources and is provided by the monitoring module of this step;

[0059] When making decisions, the system evaluates the expected utility of each topology in the topology policy library. and expected costs Expected utility The evaluation uses topology Processing level is The analytical benefits that data streams can bring; expected costs The assessment is based on the current system load. Next, execute the topology. The computing resources required;

[0060] The system is based on utility weights With cost weight Select the optimal processing topology These two weights and Similarly, the objective parameter vector of S3 Dynamic provision is used to control the system's tendency to balance value pursuit and cost savings;

[0061] To find the optimal balance between analytical value and resource consumption, the system solves a utility maximization function to determine the optimal processing topology. :

[0062] ;

[0063] in, : Indicates the selected optimal processing topology, whose data type is a topology identifier, and whose source is the calculation output of this step;

[0064] : indicates the strategy library One of the candidate topologies, which is derived from the system preset;

[0065] : Represents utility weights and cost weights, and their data type is a dimensionless floating-point number. Their origin is the objective parameter vector of S3. Provided in real time and updated dynamically;

[0066] : Represents the expected utility function, whose physical meaning is to evaluate right The analytical benefits are derived from calculations based on a pre-defined utility model;

[0067] : Represents the expected cost function, whose physical meaning is to evaluate exist The resource consumption under these conditions originates from a pre-set cost model and real-time... calculate;

[0068] To ensure the consistency of the dimensions of this formula, the expected utility Compared with expected costs They must have comparable dimensions, for example, they must all be normalized or converted into a uniform monetary value;

[0069] Selected Then, the system instantiates the data processing link according to the optimal processing topology and executes the processing on the original feature vector passed in from S1. The data processing ultimately outputs the analysis results. ;

[0070] During execution, this step also monitors and records topology reconfiguration latency. Compared with actual resource consumption These two metrics are key parameters for measuring the decision-making and execution efficiency of S2, and will be fed back to S3 for optimization.

[0071] In step S3, the system performs a post-event evaluation and closed-loop feedback on the decision-making effects of the preceding steps; the system collects the processing levels from step S1. S2 topology reconfiguration delay Actual resource consumption and analysis results ;

[0072] Based on this collected data, the system calculates the ineffective analysis resource consumption rate. and analytical performance Ineffective analysis of resource consumption rate Used to measure how much computing resources are wasted on low-value data; analytical effectiveness Used to evaluate whether the system's analytical objectives have been achieved;

[0073] The system combines ineffective analysis with resource consumption rate Topology reconfiguration delay and analytical efficiency Construct the loss function ;

[0074] Using loss function For the target parameter vector Update the target parameter vector. This is the core control parameter of this method. In this embodiment, include , , , ;

[0075] The system uses optimization algorithms such as gradient descent to... Iterative updates are performed to dynamically optimize the determination of processing level and the selection of optimal processing topology; the updated... and This will be used to determine the processing level of subsequent data streams in S1; the updated and It will be used for subsequent optimal topology selection in S2;

[0076] This invention constitutes a fully automatic and adaptive intelligent analysis method through a complete process of value-risk assessment in S1, adaptive topology reconstruction and execution in S2, and closed-loop feedback optimization in S3. It solves the fundamental problems of rigid processing flow and disconnect between resource allocation and data value in existing technologies, and realizes the dynamic and precise matching of limited computing resources to data streams with high value and high risk. While ensuring the real-time performance of key analyses, it maximizes the overall resource utilization efficiency and analysis depth of the system.

[0077] Example 2:

[0078] S1 includes:

[0079] S11. Use a pre-trained evaluation model to reason about the original feature vector and output an evaluation vector, which includes an instantaneous value component and an analytical risk component.

[0080] S12. Determine whether the risk component is greater than the risk threshold. If so, the treatment level is high risk.

[0081] S13. If the risk component is not greater than the risk threshold, then determine whether the instantaneous value component is greater than the value threshold. If so, the processing level is high value level.

[0082] S14. If the risk component analyzed is not greater than the risk threshold and the instantaneous value component is not greater than the value threshold, then the treatment level is the normal level.

[0083] Based on Example 1, this embodiment further defines the determination of the processing level in step S1. The specific logic;

[0084] In S1, the pre-trained evaluation model Output evaluation vector Including instantaneous value components and analysis of risk components Instantaneous value weight Used to quantify the potential business value of the data stream; analyze risk components. Used to quantify the probability of risk implied by the data stream;

[0085] The logic for determining the processing level is as follows:

[0086] Determine the risk components Is it greater than the risk threshold? If so, then the processing level is determined. Determined to be of high risk level ;

[0087] like Not greater than Then, the instantaneous value component is further determined. Is it greater than the value threshold? If so, then the processing level is determined. Determined to be of high value ;

[0088] like Not greater than and Not greater than Then the processing level Determined to be of standard level ;

[0089] This implementation method introduces value. and risks Two orthogonal evaluation dimensions and three processing levels enable fine-grained stratification of NFC data streams; this stratification logic ensures that the system can prioritize the allocation of the most valuable deep analysis resources. and The data stream, and for The data stream is processed at low cost, thereby maximizing the analytical value.

[0090] Example 3:

[0091] S2 includes:

[0092] S21. Extract each topology from the topology strategy library. The topology defines the combination of data preprocessing operators, the AI ​​model inference path, and the data aggregation granularity.

[0093] S22. For each processing level, evaluate the expected utility of each topology;

[0094] S23. Evaluate the expected cost of each topology for real-time system load;

[0095] S24. Combining utility weight, cost weight, expected utility, and expected cost, solve for utility maximization to select the optimal processing topology.

[0096] Based on Example 1, this embodiment further defines the optimal processing topology in step S2. Selection criteria;

[0097] Topology strategy library Each topology in All define combinations of data preprocessing operators. AI model inference path and data aggregation granularity ;

[0098] For example, one used for high-risk levels topology It may define complex feature crossover operators, heavyweight deep fraud detection models, and single-transaction granularity;

[0099] And one for regular levels topology It may only define basic cleaning operators, no AI model, and daily aggregation by merchant ID;

[0100] In selection At that time, the system performs the following evaluation:

[0101] According to the processing level Evaluate the expected utility of each topology. Expected utility It is a pre-defined utility function or lookup table; for example, the utility function can be modeled based on historical data. The value is obtained by querying a preset utility matrix. Obtain, among which Stored topology Processing level Historical average analysis benefit of the data; this historical average analysis benefit can be obtained through offline calibration, specifically by preparing labeled levels. Historical datasets, forcing the use of topology The dataset is processed, and the analysis results are evaluated; these evaluation results can be directly applied to the analytical performance defined in Example 6. As Quantization value; for the strategy library All of them and all Repeat this process at each level to construct a complete utility matrix. ;

[0102] For real-time system load Evaluate the expected cost of each topology. Expected Costs It is a dynamic function; for example, Because of Complex, its baseline cost is high, and when When the price increases, its cost... It will be dynamically increased further; a specific form of this dynamic function can be expressed as: ,in, It is topology Baseline resource cost when load is 0 It is a preset load sensitivity coefficient used to represent the cost as a function of system load. The degree of linear growth; and This can be obtained through system benchmark testing and calibration; the specific steps are as follows:

[0103] Calibration : Under system load At that time, the topology was run multiple times. Process a batch of standard data streams, measure and statistically analyze their average actual resource consumption, and use this as... ;

[0104] Calibration : Artificially setting different system loads by running disruptive tasks Run the topology at each load level And measure its actual resource consumption. Based on multiple groups Data points, using the least squares method Perform linear regression fitting and solve for the result. The value;

[0105] Combined utility weights Cost weight Expected utility and expected costs Solving for utility maximization; this process has been implemented in the embodiment of Example 1. The formula was described in detail, and the final choice was made that The largest topology as ;

[0106] This implementation abstracts complex processing flows into topology and bases them on processing levels. and system load Utility and cost By performing a dual evaluation, a quantitative decision-making process for topology selection is achieved; this ensures that the selected optimal processing topology is chosen. It can accurately match the data analysis needs and dynamically adapt to the current system resource status.

[0107] Example 4:

[0108] S2 also includes:

[0109] S25. Receive the optimal processing topology instruction, and the resource scheduler allocates computing resources, loads the required operators and AI models, and constructs the instantiated processing link;

[0110] S26. A priority queue mechanism based on processing level is adopted to route the original feature vector to the instantiated processing link;

[0111] S27. Perform the processing and output the analysis results.

[0112] Based on Example 3, this embodiment further specifies that S2 is selected The subsequent execution is the instantiation process;

[0113] Received optimal processing topology Following the instruction, the resource scheduler allocates computing resources; the resource scheduler is the core of this system's execution, and it immediately allocates computing resources according to the instructions. The definition, namely , , It requests computing resources from the resource pool and loads the required operators. and AI models Construct the instantiated processing chain;

[0114] The system employs a priority queue mechanism based on processing level to process the original feature vector input from S1. Routing to a customized, instantiated processing path; specifically, the system maintains multiple priority queues; when corresponding for or When it is in the high-priority queue, it is placed in the high-priority queue; when for When the time comes, data is placed in a low-priority queue; the instantiated processing chain will prioritize pulling data from the high-priority queue for processing.

[0115] The data is processed on the link, the processing is performed, and the analysis results are output. ;

[0116] This implementation ensures the decision-making process of S2 through the dynamic instantiation of the resource scheduler and a priority queue mechanism based on processing levels. It can be executed with low latency and high efficiency; the priority queuing mechanism ensures that high-value and high-risk data streams receive priority processing, greatly shortening the time required for critical analysis results. The output time is significantly reduced, which greatly improves the system's real-time response capability.

[0117] Example 5:

[0118] The monitoring and recording steps include:

[0119] S28. Monitor used resources and total system resources, calculate and generate system load;

[0120] S29. Record the output time of the optimal processing topology instruction, and the time when the instantiated processing link is completed and running stably;

[0121] S30. Calculate the time difference between the output time and the stable operation time to generate the topology reconfiguration delay.

[0122] During the execution of Example 4, this embodiment further defines the specific content of the monitoring and recording step in S2, which provides key feedback input for S2 and S3;

[0123] For the cost assessment of S2 Provides real-time input, and the system continuously monitors used resources. With total system resources ,pass Calculate and generate system load ;

[0124] ;

[0125] in, : Represents the system load, its data type is a normalized floating-point number of 0-1, and its source is the cost function that is calculated in real time and fed back to S2;

[0126] : Indicates the resources already used, such as the number of CPU cores, which is obtained from real-time statistics by the resource scheduler;

[0127] : Represents the total system resources, such as the total number of CPU cores, which are derived from static parameters configured in the system;

[0128] Should The real-time nature of the value ensures that S2's cost assessment can accurately reflect the current level of resource scarcity;

[0129] To quantify the overhead of topology switching, the system records the output time of the optimal topology processing command. That is, the moment when the S2 decision is completed, and the moment when the instantiated processing link is completed and running stably. That is, the moment when S2 is ready to execute;

[0130] The system calculates the time difference between these two moments to generate the topology reconfiguration delay. ;

[0131] ;

[0132] in, : Represents the topology reconstruction delay, its data type is time unit such as milliseconds, and its source is the loss function calculated and fed back to S3. ;

[0133] : Indicates a stable running time, its data type is a timestamp, and its source is recorded by the resource scheduler after the link is successfully instantiated;

[0134] : Indicates the time of instruction output; its data type is a timestamp, and its source is the output of the S2 decision system. Record in time;

[0135] It is an important negative indicator of the efficiency of the S3 evaluation system;

[0136] This embodiment, through the... and The accurate and real-time quantification provides key and measurable inputs for the cost assessment of S2 and the closed-loop optimization of S3; This ensures the current adaptability of topology selection decisions, while This ensures that the efficiency of topology switching can be evaluated and optimized.

[0137] Example 6:

[0138] S3 includes:

[0139] S31. Identify and determine the data stream set as a normal level;

[0140] S32. Obtain the actual resource consumption of the data stream at the normal level, and the minimum baseline resource cost for processing the data stream at the normal level;

[0141] S33. Calculate the difference between the actual resource consumption and the minimum baseline resource cost, and compare it with the total resource consumption of the system processing all data streams to calculate the invalid analysis resource consumption rate.

[0142] S34. By comparing the analysis results and processing levels, assess the achievement of the analysis objectives to calculate the analysis effectiveness.

[0143] Based on Example 1, this embodiment further specifies the steps in step S3 for constructing the loss function. Two key indicators and The calculation method;

[0144] Calculate the resource consumption rate of ineffective analysis :

[0145] The system identifies all systems that are classified as "normal" by S1. Data stream collection ;

[0146] Get this collection Each data stream Actual resource consumption S2 monitoring and processing of regular-level data streams at minimal baseline resource cost. ; This refers to processing The minimum cost required for data, such as archiving without analysis, is a constant preset through benchmarking;

[0147] The ineffective analysis resource consumption rate is calculated using the following formula. :

[0148] ;

[0149] in, : Represents the rate of invalid analysis resource consumption. Its data type is a floating-point number between 0 and 1, and its source is the loss function fed back to S3. ;

[0150] : Indicates the standard level The data stream set, the source of which is determined by S1;

[0151] : Indicates processing regular streams The actual consumption is monitored by the S2 execution system.

[0152] : Indicates the minimum baseline resource cost, which is derived from a preset value;

[0153] : Indicates that the system processes all data streams. The total resource consumption is monitored by S2.

[0154] It measures how much computing resources the system wastes on low-value data;

[0155] This correction is achieved by introducing... The function ensures that ineffective analysis resource consumption rate is controlled. Only positive resource waste is accumulated, avoiding the situation where... When a negative value is generated, the optimization objective of the loss function becomes clearer;

[0156] Computational analysis performance :

[0157] The system compares and analyzes the results. and processing level To assess and analyze the achievement of objectives in order to calculate analytical effectiveness. For example, for If the data, If it is successfully identified as fraud, then Increase; It is a comprehensive evaluation indicator; this comprehensive evaluation indicator It can be calculated using a specific formula, for example, in a batch. ;in, It was identified as high-risk by S1. Total number of data streams The analysis results are from this part of the data stream. This is indeed a high-risk quantity; and It is high value Statistics corresponding to the data stream; The system's accuracy in analyzing high-value and high-risk data streams was measured.

[0158] This implementation method analyzes the resource consumption rate of ineffective analyses. and analytical efficiency The precise quantification provides clear penalty and reward terms for the optimization of S3, transforming the optimization objective of S3 from the vague goal of improving efficiency into a measurable mathematical problem.

[0159] Example 7:

[0160] S3 also includes:

[0161] S35, Preset waste weight, delay weight, and performance weight;

[0162] S36. Based on waste weight, delay weight, and efficiency weight, and combined with ineffective analysis resource consumption rate, topology reconstruction delay, and analysis efficiency, construct a loss function;

[0163] S37. Calculate the gradient of the loss function with respect to the target parameter vector;

[0164] S38. Update the target parameter vector using gradients based on the preset learning rate.

[0165] Based on Example 6, this embodiment further defines the specific mathematical implementation of the closed-loop optimization in S3;

[0166] Constructing the loss function :

[0167] The system presets waste weights Delay weight With performance weight These weights are hyperparameters that are configured by the system administrator based on business objectives, such as prioritizing cost, real-time performance, or accuracy.

[0168] To eliminate dimensional inconsistencies, the system performs invalid analysis on resource consumption rates before constructing the loss function. Topology reconfiguration delay and analytical performance Normalize them so that they all fall into a uniform dimensionless interval, such as 0 to 1;

[0169] Based on these weights and the normalized metrics, a loss function is constructed. :

[0170] ;

[0171] in, : Represents the loss function value, its data type is scalar, and its source is the objective of S3 optimization, which is to minimize ;

[0172] : Represents the preset dimensionless weights, whose data type is hyperparameters, and whose source is manually configured;

[0173] : Represents the original index calculated from Examples 6 and 5. (Ineffective consumption rate) (Reconstruction Delay) and (Analysis performance) The normalized value is derived from the calculation and normalization of previous steps; for example, for topology reconstruction delay. , can be adopted The method normalizes it to the 0-1 interval, where and It is based on historical data statistics or a preset delay range; for and Additional scaling can be performed as needed;

[0174] This formula addresses three conflicting system objectives: minimizing waste. Minimize latency Maximize efficiency It has been unified into a single optimization function;

[0175] Parameter update:

[0176] System calculates loss function For the target parameter vector gradient ;Target parameter vector include ;

[0177] Based on the preset learning rate Using gradients on the target parameter vector Update; to improve parameters The robustness and convergence speed of the update process can be improved using adaptive optimization algorithms such as Adam and RMSProp; this embodiment uses gradient descent for updating.

[0178] ;

[0179] in, : Represents the target parameter vector at times t+1 and t;

[0180] : Represents the learning rate, which is a small floating-point number and is derived from preset hyperparameters, controlling the update step size;

[0181] : Represents the gradient of loss L with respect to W, its data type is a vector, and its source is calculated through backpropagation or numerical methods; given that the threshold determination in S1 and the optimal topology selection in S2 are inherently discrete and non-differentiable, the loss function... Gradients cannot be directly calculated using standard end-to-end backpropagation. Those skilled in the art can solve this problem in at least one of the following ways to optimize S3:

[0182] A perturbation-based numerical estimation method is adopted;

[0183] For example, using a synchronous perturbation stochastic approximation algorithm; at time t, for the current parameter vector Apply a small, randomly generated perturbation vector Calculate the system respectively and Loss at two points and ;pass To estimate right No. The partial derivatives of each component are used to construct the gradient vector. ;

[0184] A policy gradient method based on reinforcement learning is adopted;

[0185] target parameter vector Consider it as a strategy that determines the decision-making behavior in S1 and S2; calculate within a batch This is considered a reward obtained through this strategy; The update objective is to maximize this reward; since the decisions in S1 and S2 are discrete, those skilled in the art can use the REINFORCE algorithm or its variants to estimate the policy gradient. This bypasses the problem of non-differentiability; for example, randomness can be introduced into the threshold determination to construct a stochastic policy, and then the policy gradient theorem can be applied for updating.

[0186] This update rule makes The four key decision parameters can automatically and iteratively move towards maximizing the total loss. Minimal directional adjustment;

[0187] This implementation constructs a comprehensive loss function that includes waste, delay, and efficiency. And update the objective parameter vector using gradient descent. This enables the control of the system's core decision parameters. Automatic, closed-loop, and intelligent optimization.

[0188] Example 8:

[0189] The original transaction data stream includes product code, weight, unit price, timestamp, and merchant ID. The original feature vector is constructed based on the extraction from the original transaction data stream.

[0190] This implementation method further defines step S1 in Example 1, clarifying the data basis for AI analysis;

[0191] The raw transaction data stream includes product code, weight, unit price, timestamp, and merchant ID;

[0192] Original feature vector It is built upon the extraction of these raw transaction data streams; the pre-trained evaluation model in S1. Instead of directly processing the raw data stream, it processes the raw feature vectors that have undergone feature engineering. ;For example:

[0193] Merchant ID can be used to extract features such as the merchant's historical transaction frequency, average order value, and historical risk records.

[0194] Commodity codes can be associated with characteristics such as the category of the commodity, its perishability, and whether it is a high-value commodity;

[0195] Timestamps can be used to extract time-series characteristics such as when a transaction occurred and whether it occurred on a weekday.

[0196] Weight and unit price can be normalized, and the total amount can be calculated as a feature;

[0197] After undergoing the aforementioned feature engineering and transformations, these raw data are used to construct high-dimensional vectors. Only then can it be evaluated by the model in S1. Effectively used for value and risks Assessment;

[0198] This implementation method clarifies the raw data foundation for AI analysis; by utilizing this multi-dimensional raw data, a pre-trained evaluation model is developed. It can more comprehensively and accurately assess the instantaneous value of data streams. and risk analysis This provides a high-quality basis for subsequent refined stratification and topology selection.

[0199] Example 9:

[0200] After the target parameter vector is updated, the updated value threshold and risk threshold are used for subsequent determination of the processing level in S1; the updated utility weight and cost weight are used for subsequent selection of the optimal processing topology in S2.

[0201] Based on Example 7, this implementation clarifies the feedback path and function of S3 parameter update, forming a complete adaptive closed loop;

[0202] When the target parameter vector Updated in S3 After being calculated:

[0203] Updated value threshold and risk threshold Right now The corresponding component in the data stream will be pushed to the processing level determination module in step S1; when processing the next batch of NFC data streams, S1 will be used for subsequent determination of the processing level.

[0204] Updated utility weights and cost weight Right now The corresponding component in the process will be pushed to the optimal processing topology selection module in step S2; when selecting a topology for subsequent data streams in S2, it will be used for subsequent selection of the optimal processing topology.

[0205] This implementation method clarifies the output of S3. How to feed back to S1 to have an impact? And S2 to influence This constructs a complete and dynamic closed loop of evaluation-decision-execution-optimization; this enables the method to break free from the dependence on static thresholds and weights, and to learn autonomously and converge to the optimal strategy, ensuring that the system always maintains near-optimal resource allocation efficiency and analytical performance under changing business needs and system loads.

[0206] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI, characterized in that, The specific steps include: S1. Collect the raw transaction data stream of the NFC electronic price-computing scale in real time, extract key features, and construct the raw feature vector; use a pre-trained evaluation model to infer the raw feature vector and output the evaluation vector; preset value threshold and risk threshold, compare the evaluation vector with the value threshold and risk threshold to obtain the processing level; S2. Based on the processing level and combined with the preset topology strategy library and real-time system load, evaluate the expected utility and expected cost of each topology in the topology strategy library; select the optimal processing topology according to the utility weight and cost weight; instantiate the data processing link according to the optimal processing topology, perform data processing of the original feature vector, and output the analysis results; monitor and record the topology reconstruction delay and actual resource consumption, as well as the analysis results. S3. Collect processing level and topology reconstruction delay, and collect actual resource consumption and analysis results; calculate ineffective analysis resource consumption rate and analysis efficiency; combine ineffective analysis resource consumption rate, topology reconstruction delay, and analysis efficiency to construct a loss function; use the loss function to update the target parameter vector, which includes value threshold and risk threshold, as well as utility weight and cost weight, to dynamically optimize the determination of processing level and the selection of optimal processing topology; S3 includes: S31. Identify and determine the data stream set as a normal level; S32. Obtain the actual resource consumption of the data stream at the normal level, and the minimum baseline resource cost for processing the data stream at the normal level; S33. Calculate the difference between the actual resource consumption and the minimum baseline resource cost, and compare it with the total resource consumption of the system processing all data streams to calculate the invalid analysis resource consumption rate. S34. By comparing the analysis results and processing levels, assess the achievement of the analysis objectives to calculate the analysis effectiveness.

2. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 1, characterized in that, S1 includes: S11. Use a pre-trained evaluation model to reason about the original feature vector and output an evaluation vector, which includes an instantaneous value component and an analytical risk component. S12. Determine whether the risk component is greater than the risk threshold. If so, the treatment level is high risk. S13. If the risk component is not greater than the risk threshold, then determine whether the instantaneous value component is greater than the value threshold. If so, the processing level is high value level. S14. If the risk component analyzed is not greater than the risk threshold and the instantaneous value component is not greater than the value threshold, then the treatment level is the normal level.

3. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 1, characterized in that, S2 include: S21. Extract each topology from the topology strategy library. The topology defines the combination of data preprocessing operators, the AI ​​model inference path, and the data aggregation granularity. S22. For each processing level, evaluate the expected utility of each topology; S23. Evaluate the expected cost of each topology for real-time system load; S24. Combining utility weight, cost weight, expected utility, and expected cost, solve for utility maximization to select the optimal processing topology.

4. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 3, characterized in that, S2 also includes: S25. Receive the optimal processing topology instruction, and the resource scheduler allocates computing resources, loads the required operators and AI models, and constructs the instantiated processing link; S26. A priority queue mechanism based on processing level is adopted to route the original feature vector to the instantiated processing link; S27. Perform the processing and output the analysis results.

5. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 4, characterized in that, The monitoring and recording steps include: S28. Monitor used resources and total system resources, calculate and generate system load; S29. Record the output time of the optimal processing topology instruction, and the time when the instantiated processing link is completed and running stably; S30. Calculate the time difference between the output time and the stable operation time to generate the topology reconfiguration delay.

6. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 1, characterized in that, S3 also includes: S35, Preset waste weight, delay weight, and performance weight; S36. Based on waste weight, delay weight, and efficiency weight, and combined with ineffective analysis resource consumption rate, topology reconstruction delay, and analysis efficiency, construct a loss function; S37. Calculate the gradient of the loss function with respect to the target parameter vector; S38. Update the target parameter vector using gradients based on the preset learning rate.

7. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 1, characterized in that, The original transaction data stream includes product code, weight, unit price, timestamp, and merchant ID. The original feature vector is constructed based on the extraction from the original transaction data stream.

8. The method for big data acquisition, processing, and intelligent analysis of electronic price-computing scales based on NFC and AI according to claim 6, characterized in that, After the target parameter vector is updated, the updated value threshold and risk threshold are used for subsequent determination of the processing level in S1; the updated utility weight and cost weight are used for subsequent selection of the optimal processing topology in S2.

Images