A business data real-time analysis processing method based on a lightweight neural network

Abstract

The application discloses a kind of based on lightweight neural network's service data real-time analysis processing method, belong to data processing technical field.This method includes: obtaining real-time service data stream and completing intelligent preprocessing and redundancy optimization;Lightweight compression is carried out to initial neural network using structured pruning, knowledge distillation and mixed precision quantification multi-strategy fusion mode;Lightweight model is divided into end side, side, cloud side submodel and is deployed hierarchically;End side executes first level real-time inference, and according to the comparison result of inference confidence and dynamic optimization threshold, hierarchical scheduling edge second level inference or cloud third level inference;While continuously monitoring service data distribution change, identifying concept drift and using multi-model replica weighted average realizes online incremental adaptive update.The application realizes the intelligent real-time analysis processing of low delay, high accuracy, self-adaptive, iterative of service data, effectively improves end edge cloud collaborative inference efficiency and model long-term adaptation ability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method for real-time analysis and processing of business data based on lightweight neural networks. Background Technology

[0002] With the rapid popularization of digital business scenarios, streaming data generated by various business systems is characterized by massive volume, high concurrency, strong real-time requirements, and rapid dynamic changes. Intelligent real-time analysis and processing has become a core supporting technology for business operations, risk assessment, and data-driven decision-making. Against the backdrop of industry development towards real-time big data processing, edge intelligent inference, and closed-loop iterative optimization of business data, the technical limitations of traditional business data analysis and processing solutions are becoming increasingly apparent, making it difficult to meet the demands for intelligent business analysis that requires high real-time performance, high precision, low power consumption, and adaptive iteration.

[0003] Existing real-time business data analysis solutions mostly adopt a centralized cloud-based big data processing architecture, relying on traditional heavy neural network models to complete data feature extraction and intelligent analysis. These models have a large number of parameters, high computational redundancy, and extremely high requirements for device computing power, memory, and power consumption, making them unsuitable for low-resource deployment scenarios on terminals and edge devices. Traditional processing architectures suffer from high inference latency, high computing resource consumption, and the inability to perform local real-time inference, making it difficult to support the millisecond-level real-time analysis requirements of high-concurrency business data streams.

[0004] Meanwhile, existing neural network models employ a single inference mode, mostly using purely cloud-based centralized inference or purely edge-based local inference, without establishing a layered collaborative inference architecture between the edge and cloud. Purely edge-based inference models lack accuracy and fault tolerance, making them highly susceptible to biased analysis results; purely cloud-based inference relies on stable communication links, and is prone to inference lag, data latency, and task interruption in scenarios with network fluctuations and link congestion, failing to balance inference real-time performance and analysis accuracy. Traditional inference solutions use fixed threshold judgment mechanisms, which cannot dynamically adjust the judgment criteria according to the accuracy and latency requirements of business scenarios, resulting in poor scenario adaptability and a tendency for misjudgments and omissions in analysis.

[0005] On the other hand, existing business data analysis models generally adopt static training and offline update modes, with fixed model parameters, making it impossible to perceive changes in the distribution of business data streams in real time. In actual business operations, affected by factors such as business iteration, scenario updates, changes in user behavior, and fluctuations in system operating status, business data streams are prone to conceptual drift, leading to decreased adaptability of static models and continuous degradation of analytical accuracy. Traditional solutions lack effective online incremental updates and adaptive iteration mechanisms, making it impossible to achieve continuous optimization of model performance.

[0006] Furthermore, traditional business data preprocessing methods are limited, only capable of basic data cleaning and format standardization, and do not perform redundant optimization for video business data or high-frequency dynamic data streams, resulting in low data processing efficiency. Conventional neural network lightweighting methods are also limited, employing only simple pruning or quantization, leading to significant accuracy loss and weak generalization ability after model compression. Moreover, they do not incorporate multi-level resource constraints at the edge, cloud, and device levels for hierarchical model segmentation and deployment, resulting in low resource utilization and overall inference efficiency.

[0007] In summary, existing real-time business data analysis technologies suffer from several technical shortcomings, including bloated models, high inference latency, simplistic inference architecture, poor scenario adaptability, lack of dynamic threshold control mechanisms, inability to adapt to data concept drift, weak model iteration capabilities, low data preprocessing efficiency, and insufficient utilization of edge-cloud resources. There is an urgent need to develop a lightweight, layered, collaborative, dynamically adaptive, and closed-loop iterative optimization solution for real-time business data analysis. Summary of the Invention

[0008] The purpose of this invention is to provide a real-time analysis and processing method for business data based on lightweight neural networks, in order to solve the technical problems of traditional business data analysis models being bloated and inefficient, having a single inference architecture, poor adaptability to fixed thresholds, being unable to adapt to data concept drift, being unable to iterate online, having low data preprocessing accuracy, and insufficient utilization of edge-cloud resources. This invention aims to achieve intelligent real-time analysis and processing of business data with low latency, high accuracy, adaptability, and iterative capability.

[0009] This invention adopts the following technical solution: a real-time analysis and processing method for business data based on lightweight neural networks, comprising the following steps: S1. Data Acquisition: Acquire real-time business data streams generated by the business system, and complete the real-time collection and access of full-volume streaming business data.

[0010] S2. Data Preprocessing: The acquired real-time business data stream is preprocessed by data cleaning, format unification, and timestamp alignment; at the same time, for video business data, motion gating frame selection technology is used to filter redundant frames, retaining only the frame data where the screen content changes; according to the data change rate of the real-time business data stream, the data sampling frequency is dynamically adjusted through adaptive sampling frequency adjustment technology to achieve redundant data optimization processing.

[0011] S3. Model Lightweighting: At least one lightweighting strategy among structured pruning, quantization, and knowledge distillation is used to compress the initial neural network model to obtain a lightweight neural network model. Specifically, redundant convolutional filters are removed through structured pruning based on the L2 norm of the filter. The model's feature extraction capability is optimized by combining feature-level and logit-level joint knowledge distillation training. Finally, the model weights and activation values ​​are compressed through mixed precision quantization. At the same time, hierarchical dynamic pruning threshold adjustment is achieved based on the median absolute deviation method to ensure a balance between model compression accuracy and inference efficiency.

[0012] S4. Layered Model Segmentation and Deployment: According to the preset segmentation strategy, combined with the constraints of computing power, memory, power consumption, and bandwidth resources of terminal devices, edge servers, and cloud servers, as well as the requirements for business data inference latency tolerance and data privacy protection level, the lightweight neural network model is segmented into terminal-side sub-models, edge-side sub-models, and cloud-side sub-models; wherein the terminal-side sub-model retains the shallow feature extraction structure, the edge-side sub-model retains the mid-level feature extraction structure, and the cloud-side model retains the deep feature extraction and task inference structure, and is deployed on terminal devices, edge servers, and cloud servers respectively.

[0013] S5. Terminal-side first-level real-time inference: The preprocessed real-time business data stream is input into the terminal-side sub-model to perform first-level inference, obtain the first inference result and the corresponding confidence score, and realize localized rapid preliminary analysis of business data.

[0014] S6. Hierarchical inference scheduling: The confidence score is compared with a preset threshold, and the corresponding inference task scheduling logic is executed according to the comparison result; wherein the preset threshold can be dynamically solved for the optimal value through a multi-objective optimization algorithm according to the accuracy and latency requirements of the business scenario, and adjusted in real time as the business scenario changes.

[0015] If the confidence score is greater than or equal to the preset threshold, the first inference result is directly output and used as the final analysis result. If the confidence score is less than the preset threshold, the real-time business data stream or the intermediate feature data output by the edge sub-model is uploaded to the edge server, and the edge sub-model performs the second-level inference to obtain the second inference result. If the confidence score of the second inference result is still less than the preset threshold, the intermediate feature data or the high-dimensional feature data output by the side sub-model is uploaded to the cloud server, the third-level inference is performed through the cloud side sub-model, the third inference result is output and used as the final analysis result.

[0016] S7. Online Adaptive Update: During real-time inference, continuously monitor the data distribution changes of the real-time business data stream; collect the current business data feature distribution based on a sliding time window, calculate the difference measure with the historical feature distribution, and when a concept drift in the data distribution is detected, use an online weighted averaging method, combined with the inference performance of multiple time window model replicas, to update the weight parameters of the lightweight neural network model, thereby realizing online incremental adaptive iteration of the model.

[0017] This invention also provides a real-time business data analysis and processing system based on lightweight neural networks, used to implement the aforementioned analysis and processing methods. The system constructs a three-layer collaborative analysis architecture of terminal, edge, and cloud, specifically including: The data acquisition module is used to acquire real-time business data streams generated by the business system; The data preprocessing module is used to perform data cleaning, format unification, timestamp alignment, redundant frame filtering, and adaptive sampling processing on the real-time business data stream. The model compression module is used to perform lightweight compression of the initial neural network model through a combination of strategies including structured pruning, knowledge distillation, and mixed precision quantization. The model segmentation and deployment module is used to segment lightweight neural network models based on multi-dimensional constraints and to deploy sub-models on the edge, cloud, and device sides. The edge-side inference engine, deployed on the terminal device, is used to perform the first-level real-time inference and output the first inference result and the corresponding confidence score; The hierarchical scheduling module is used to schedule inference tasks at different levels based on the comparison results of the confidence scores of the inference results with preset thresholds. The edge inference engine, deployed on edge servers, is used to perform second-level inference tasks; The cloud-based inference engine, deployed on a cloud server, is used to perform Level 3 inference tasks and output the final analysis results. The online adaptive update module is used to detect the concept drift phenomenon in business data streams and complete the online incremental update of the lightweight neural network model based on real-time business data.

[0018] The core lightweight and adaptive iteration principles of this solution are as follows: The proposed solution employs L2 norm filtering to quantify the contribution of convolutional layer weights, and uses hierarchical dynamic thresholding to perform structured pruning, accurately eliminating redundant parameters and significantly simplifying the model structure while preserving core feature extraction capabilities. Combined with a teacher-student network knowledge distillation mechanism, the feature information and classification logic of the high-precision initial model are transferred to the simplified model, compensating for the accuracy loss caused by pruning. Finally, hybrid precision quantization is used to compress model storage and computational overhead, achieving a balance between lightweight and high accuracy.

[0019] This invention constructs a hierarchical inference architecture across edge and cloud, allocating inference tasks based on the resource characteristics of devices at different levels. It achieves cross-level data interaction through small amounts of irreversible intermediate feature data, reducing communication overhead while ensuring the privacy and security of business data. Furthermore, it dynamically optimizes the inference confidence threshold based on the accuracy and latency constraints of different business scenarios, achieving adaptive adaptation to various business scenarios.

[0020] This invention monitors changes in business data distribution in real time through a sliding time window, accurately identifies concept drift issues caused by business scenario iterations, and achieves smooth iterative optimization of model parameters through an online update mechanism of weighted average of multiple model replicas. This avoids model lag and analysis accuracy degradation caused by offline updates, forming a closed-loop optimization system of real-time analysis, real-time monitoring, and real-time iteration.

[0021] The beneficial effects achieved by the present invention using the above technical solution are as follows: 1. Employing at least one lightweight strategy among structured pruning, quantization, and knowledge distillation, and preferably using a multi-strategy fusion approach, combined with a hierarchical dynamic pruning threshold mechanism, it can accurately eliminate redundant model parameters, significantly reduce the number of model parameters and computational overhead, and adapt to low-computing-power, low-memory, and low-power deployment scenarios of terminals and edge devices. It effectively solves the problems of high inference latency, high resource consumption, and inability to be deployed locally at the edge in traditional heavy neural network models, and significantly improves the efficiency of real-time analysis of business data.

[0022] 2. By constructing a layered collaborative and hierarchical inference architecture for edge, cloud, and end devices, the model is divided into shallow, medium, and deep feature inference structures and deployed on edge, cloud, and end devices respectively. Tasks are dynamically scheduled based on inference confidence, taking into account the low latency advantage of edge inference and the high accuracy advantage of cloud inference. At the same time, through intermediate feature data interaction, the amount of data transmission is reduced, adapting to the business analysis needs under complex network conditions.

[0023] 3. By implementing scenario-based dynamic adaptive adjustment of the inference threshold, the optimal threshold is solved through a multi-objective optimization algorithm based on the accuracy and latency requirements of the business scenario. This can adapt to the analysis and judgment requirements of different business scenarios and different operating states, effectively avoiding the analysis misjudgment and omission problems caused by traditional fixed thresholds, and greatly improving the accuracy of business data analysis and scenario adaptability.

[0024] 4. By segmenting the lightweight neural network model according to the feature extraction hierarchy and taking into account the constraints of computing power, memory, power consumption and bandwidth resources of terminal, edge and cloud devices for differentiated deployment, a layered collaborative inference architecture with fast response on the edge, flexible expansion at the edge and deep analysis in the cloud is achieved. At the same time, only intermediate feature data is transmitted across layers instead of the original business data, which effectively reduces communication overhead, ensures the privacy and security of business data, and improves the resource utilization efficiency and deployment flexibility of the overall system.

[0025] 5. By establishing a real-time monitoring and online incremental update mechanism for business data concept drift, changes in business data distribution can be perceived in real time. The model can be iterated online through a weighted average of multiple model replicas without downtime for training and updates. This allows for continuous adaptation to iterative changes in business scenarios, effectively solving the problem of accuracy decay in static models over long-term operation and achieving closed-loop self-optimization and upgrade of the analysis model.

[0026] 6. By optimizing the business data preprocessing mechanism, redundant frame filtering and adaptive frequency sampling technology are used to remove invalid and redundant data, thereby improving data processing efficiency and the proportion of effective data, further reducing the pressure of subsequent inference calculations, and improving the real-time performance and stability of overall business analysis and processing. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a schematic diagram of the edge-cloud hierarchical model segmentation deployment and hierarchical inference architecture of the present invention; Figure 3 This is a schematic diagram of the lightweight neural network processing flow of the present invention; Figure 4 This is a schematic diagram of the business data concept drift detection and online adaptive update process of the present invention; Figure 5 This is a schematic diagram of the dynamic threshold optimization and hierarchical reasoning scheduling logic of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, the present invention provides a real-time business data analysis and processing method based on a lightweight neural network, comprising a processing flow of steps S1 to S7. Through a closed-loop architecture of real-time data access, intelligent preprocessing, lightweight model compression, edge-cloud layered deployment, multi-level collaborative inference, dynamic threshold scheduling, and online adaptive iteration, it achieves real-time intelligent analysis of massive streaming business data with low latency, high precision, strong adaptability, and high privacy and security. This solves the technical problems of traditional business data analysis models being bloated, having lagging inference, poor scenario adaptability, inability to cope with data concept drift, and lagging model iteration.

[0030] S1. Real-time business data stream full-domain acquisition and access: Captures streaming business data continuously generated during the operation of various business systems in real time, completes real-time access, time-series aggregation and raw caching of the full business data stream, covering multiple types of heterogeneous business data such as business operation data, equipment operation data, business status logs, and video surveillance data, ensuring the data integrity and real-time performance of subsequent analysis, and providing the raw data foundation for subsequent preprocessing, model inference and status assessment.

[0031] S2. Multi-dimensional intelligent data preprocessing and redundancy optimization: Standardized preprocessing operations are uniformly performed on the real-time business data streams accessed in step S1, including data cleaning, abnormal noise removal, format unification, and time-series timestamp alignment, eliminating format differences, time-series misalignments, and invalid interference data from multi-source heterogeneous data. Simultaneously, relying on the redundancy optimization mechanism configured in step S2, a motion-gated frame selection strategy is adopted for video business data streams, filtering static, unchanging redundant frames and retaining only valid changing frames. The formula for calculating inter-frame differences is as follows: In the formula, This refers to the pixel difference between video frames. Let be the grayscale value of the pixel corresponding to the video frame at time t. This represents the grayscale value of the pixel corresponding to the previous video frame. These represent the height and width in pixels of the video frame, respectively; when the inter-frame difference... If the current frame is less than the preset static threshold, it is determined to be a redundant static frame and filtered out.

[0032] For high-frequency fluctuating business data streams, the sampling frequency is adaptively adjusted based on the data change rate. The variance of business data within a sliding window is calculated in real time to characterize the severity of data fluctuations. The optimal sampling frequency is adaptively matched to reduce the amount of data computation for subsequent model inference while ensuring the integrity of business data features, and outputs a high-quality standardized business data stream.

[0033] S3. Multi-strategy fusion neural network lightweight compression processing: Based on the lightweighting process in step S3, multi-level lightweight compression optimization, including structured pruning, joint knowledge distillation, and mixed-precision quantization, is performed on the initial high-precision neural network model. First, the contribution of each convolutional layer parameter is quantified based on the L2 norm of the filter. The formula for calculating the L2 norm of a single filter is as follows: In the formula, The L2 norm of a single convolutional filter. These are the weight parameters inside the filter. This represents the total number of filter parameters; the smaller the L2 norm value, the weaker the feature extraction capability and the higher the parameter redundancy of the corresponding filter.

[0034] This invention combines the median absolute deviation mechanism to configure an independent dynamic pruning threshold for each network layer. The calculation formula for the dynamic pruning threshold of a single layer is as follows: In the formula, The dynamic pruning threshold for a single network layer. This is the median of the L2 norm of all filters in this layer. This is the pruning adjustment coefficient. This is the median absolute deviation of the L2 norm of the filter layer, used to characterize the discrete characteristics of the parameters, and to accurately eliminate redundant filters based on the independent dynamic thresholds of each layer.

[0035] Then, through joint knowledge distillation at the feature level and logit level, the high-precision feature extraction capability and inference logic of the initial teacher network are transferred to the pruned and simplified student network, compensating for the accuracy loss during pruning. The formula for the total loss function of joint distillation is as follows: In the formula, This represents the total loss during distillation. For characteristic layer distillation loss, For the output layer logit distillation loss, This is the weight balancing coefficient, used to balance the weight ratio between feature fitting and result fitting. Finally, mixed precision quantization is used to compress the model weights and activation values, significantly reducing the number of model parameters and computational overhead while keeping accuracy loss under control. This results in a lightweight neural network model that balances lightweight design, high precision, and high inference speed, adaptable to differentiated deployment scenarios across multiple levels, including edge, cloud, and device.

[0036] S4. Layered Model Segmentation and Differentiated Deployment: Based on the layered deployment strategy in step S4, and considering the constraints of computing power, memory, power consumption, and bandwidth resources of terminal, edge, and cloud devices, combined with the requirements of business inference latency tolerance and data privacy protection level, the lightweight neural network model obtained in step S3 is structurally segmented. The shallow feature extraction structure of the model is segmented into an edge sub-model, the mid-layer feature fusion structure is segmented into an edge sub-model, and the deep feature and task inference structure is segmented into a cloud sub-model. These sub-models are deployed on terminal devices, edge servers, and cloud servers, respectively, forming a layered collaborative deployment architecture of initial inference at the edge, fine-grained correction at the edge, and high-precision judgment at the cloud.

[0037] S5. Real-time fast inference at the first level on the edge: Based on the first-level inference process in step S5, the standardized real-time business data stream preprocessed in step S2 is input into the deployed edge sub-model, shallow feature extraction and first-level real-time inference are performed, and preliminary business analysis results and corresponding inference confidence scores are quickly output. This enables localized millisecond-level rapid analysis of business data and meets the low-latency analysis requirements of conventional business scenarios.

[0038] S6. Confidence Dynamic Optimization and Multi-Level Inference Intelligent Scheduling: Based on the hierarchical scheduling mechanism in step S6, a multi-objective optimization model is constructed based on the dual constraints of accuracy and latency in the business scenario. The optimal confidence threshold is dynamically solved and updated adaptively in real time. The formula for the multi-objective optimization model is as follows: In the formula, Preset a threshold for confidence level. Let the inference accuracy function be . For inference delay function, These represent the upper and lower limits of the threshold, respectively. The optimal threshold suitable for the current business scenario is obtained through multi-objective optimization.

[0039] The first-level inference confidence output in step S5 is compared with the dynamic optimal threshold, and different inference tasks are triggered in stages: if the confidence meets the standard, the edge inference result is directly output as the final analysis result; if the confidence does not meet the standard, the intermediate feature data is uploaded to the edge server to perform second-level inference; if the edge inference confidence still does not meet the requirements, it is further uploaded to the cloud to perform third-level high-precision inference, realizing a smart hierarchical scheduling logic of low latency priority and high precision fallback.

[0040] S7. Data Concept Drift Detection and Online Adaptive Model Update: Based on the online iteration mechanism of step S7, real-time business data feature distribution is collected in real time through a sliding time window. The KL divergence metric is used to measure the distribution difference between real-time data and historical steady-state data, accurately identifying the concept drift phenomenon in the business data stream. The KL divergence distribution difference measurement formula is as follows: In the formula, The KL divergence value between the real-time data distribution and the historical steady-state data distribution. This represents the characteristic probability distribution of real-time business data within the current sliding window. This represents the standard characteristic probability distribution of historical steady-state business data; the larger the KL divergence value, the more significant the data distribution shift and the higher the degree of concept drift.

[0041] When a data distribution offset is detected to be excessive, the model weight parameters are updated iteratively based on the actual inference performance of the multi-time-window model replicas. The model weight update formula is as follows: In the formula, These are the model weight parameters after iterative updates. The model replica weights for each time window. The performance weight coefficients for corresponding model replicas are set, with higher weight coefficients for model replicas that have better inference performance. This weighted update method enables online incremental adaptive optimization of lightweight neural network models without downtime or interruption, continuously adapting to business scenario iterations and data distribution changes, and maintaining optimal inference performance of the model in the long term.

[0042] like Figure 2 As shown, the real-time business data analysis and processing system of this invention adopts a layered, modular, and decoupled architecture of edge, cloud, and endpoint. Each module operates independently, collaborates, and forms a closed-loop function. Specifically, it includes: a data acquisition module, a data preprocessing module, a model compression module, a model segmentation and deployment module, an endpoint inference engine, a hierarchical scheduling module, an edge inference engine, a cloud inference engine, and an online adaptive update module. Among them, the data acquisition module is responsible for the continuous acquisition and access of real-time business data streams across the entire domain; the data preprocessing module is responsible for the standardization processing of multi-source data and intelligent optimization of redundant data; the model compression module is responsible for the multi-level lightweight compression of the initial neural network model; the model segmentation and deployment module is responsible for the hierarchical segmentation of the lightweight model and the deployment of multi-level devices; the endpoint inference engine, edge inference engine, and cloud inference engine complete the first, second, and third-level collaborative inference step by step; the hierarchical scheduling module is responsible for dynamic threshold optimization and intelligent scheduling of multi-level inference tasks; and the online adaptive update module is responsible for data drift detection and online iterative optimization of the model, thus constructing a closed-loop system for intelligent business data analysis throughout the entire process.

[0043] like Figure 3 As shown, the specific process of model lightweighting in step S3 of this invention is as follows: First, a high-precision initial neural network model trained offline is obtained. The redundancy of parameters in each convolutional layer is calculated using the L2 norm of the filter. The independent dynamic pruning threshold for each network layer is generated by combining the median absolute deviation algorithm to complete the structured pruning and remove redundant parameters. Using the initial model as the teacher network and the pruned model as the student network, joint knowledge distillation training at the feature level and logit level is performed to complete the transfer of feature capabilities and inference logic. Finally, the model weights and activation values ​​are compressed by mixed precision quantization to output the final lightweight neural network model, achieving comprehensive optimization of model size, computational load, and inference latency.

[0044] like Figure 4 As shown, the specific process of online adaptive update in step S7 of this invention is as follows: The system continuously collects real-time business data through a sliding time window, and constructs real-time data feature distribution and historical steady-state feature distribution respectively; the distribution difference is calculated through KL divergence to determine whether concept drift has occurred; if data drift is detected, the pre-maintained multi-time window model replicas are retrieved, and the model weights are updated based on the real-time inference performance of each replica; the updated new model is put into real-time inference business, and the monitoring and updating are carried out iteratively to continuously adapt to the dynamic changes of business data.

[0045] like Figure 5 As shown, the specific logic of step S6, dynamic threshold optimization and hierarchical inference scheduling, of this invention is as follows: A multi-objective optimization function is constructed based on the real-time accuracy and latency requirements of the current business scenario. The real-time optimal confidence threshold is obtained through Pareto optimal solution. The edge-side first-level inference confidence is compared with the dynamic threshold. High-confidence results are directly output, while low-confidence results are progressively upgraded to the edge and cloud to complete multi-level inference. The threshold parameters are dynamically updated throughout the process to adapt to the scheduling requirements of different business scenarios and different data states, achieving a dynamic balance between real-time performance and accuracy.

[0046] Performance comparison test and effect verification: To objectively and comprehensively verify the practical application performance of the technical solution of this invention, this embodiment conducts a parallel comparative experiment under the same business dataset, the same operating environment, and the same hardware equipment conditions, using a traditional fixed model and single-level inference business data analysis scheme as a control. The measured results of each performance index are shown in the table below: As can be seen from the above experimental data, compared with traditional business data analysis solutions, this invention significantly reduces model inference latency and resource consumption through multi-strategy model lightweighting, edge-cloud layered collaborative inference, dynamic threshold control, concept drift adaptive iteration, and data preprocessing optimization technologies. It also significantly improves the accuracy of business data analysis and the adaptability to complex scenarios, enabling efficient, accurate, and intelligent analysis and processing of various real-time business data. It has strong engineering applicability and industry universality.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0048] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for real-time analysis and processing of business data based on lightweight neural networks, characterized in that, Includes the following steps, S1. Data Acquisition: Acquire real-time business data streams generated by the business system; S2. Data preprocessing: Perform data cleaning, format unification, and timestamp alignment preprocessing on the acquired real-time business data stream; S3. Model Lightweighting: At least one lightweighting strategy among structured pruning, quantization, and knowledge distillation is used to compress the initial neural network model to obtain a lightweight neural network model. S4. Layered model segmentation and deployment: According to the preset segmentation strategy, the lightweight neural network model is segmented into terminal sub-model, edge sub-model and cloud sub-model, and the terminal sub-model is deployed on the terminal device, the edge sub-model is deployed on the edge server and the cloud sub-model is deployed on the cloud server. S5. Terminal-side first-level real-time inference: Input the preprocessed real-time business data stream into the terminal-side sub-model, perform first-level inference, and obtain the first inference result and the corresponding confidence score. S6. Hierarchical Inference Scheduling: The confidence score is compared with a preset threshold, and the corresponding inference task scheduling logic is executed based on the comparison result. If the confidence score is greater than or equal to the preset threshold, the first inference result is directly output and used as the final analysis result. If the confidence score is less than the preset threshold, the real-time business data stream or the intermediate feature data output by the edge sub-model is uploaded to the edge server, and the edge sub-model performs the second-level inference to obtain the second inference result. If the confidence score of the second inference result is still less than the preset threshold, the intermediate feature data or the high-dimensional feature data output by the side sub-model is uploaded to the cloud server, the third-level inference is performed through the cloud side sub-model, the third inference result is output and used as the final analysis result; S7. Online Adaptive Update: During real-time inference, continuously monitor the data distribution changes of the real-time business data stream; when a concept drift is detected in the data distribution, perform online incremental updates to the lightweight neural network model based on the current real-time business data stream.

2. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, The S3 model lightweighting steps specifically include: S301, Structured pruning: A structured pruning strategy based on the L2 norm of filters is adopted. The L2 norm of each convolutional filter in the initial neural network model is calculated. Based on a preset percentile threshold, filters with L2 norms lower than the threshold are pruned to obtain the pruned network structure. S302, Knowledge Distillation Training: Using the initial neural network model as the teacher network and the pruned network as the student network, the student network is trained by combining feature-level distillation and logit-level distillation. S303, Hybrid Precision Quantization: The weights and activation values ​​of the pruned and distilled network that have been trained are quantized and compressed to obtain the final lightweight neural network model.

3. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 2, characterized in that, The structured pruning in step S301 employs a dynamic threshold-driven mechanism, specifically including: Based on the median absolute deviation method, the dynamic pruning threshold corresponding to the L2 norm of each layer filter in the initial neural network model is calculated. Based on the distribution characteristics of the L2 norm of the filters in each network layer, independent pruning thresholds are configured for different network layers.

4. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, The segmentation strategy in step S4 is determined by multiple constraints, including: computing resource constraints, memory resource constraints, and power consumption constraints of the terminal device; computing resource constraints and communication bandwidth constraints of the edge server; computing resource constraints of the cloud server; inference latency tolerance requirements of the real-time business data stream; and data privacy protection level requirements of the real-time business data stream.

5. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, The preset threshold in step S6 can be dynamically adjusted according to the business scenario. Specific adjustment methods include: Obtain the accuracy and latency requirements for the current business scenario; Based on the aforementioned accuracy and latency requirements, the optimal value of the preset threshold is determined using a multi-objective optimization algorithm. During real-time inference execution, the preset threshold is dynamically adjusted according to real-time changes in the business scenario.

6. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, In step S7, the method for detecting the conceptual drift of data distribution includes: The feature distribution of the current real-time business data stream is collected based on a sliding time window. Calculate the distribution difference measure between the current feature distribution and the historical feature distribution; When the distribution difference metric is greater than a preset drift detection threshold, it is determined that the real-time business data stream has experienced concept drift.

7. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, The online incremental update in step S7 adopts an online weighted averaging method, and the specific steps include: During the execution of the real-time inference step, multiple copies of the neural network model trained under different time windows are maintained; Statistical analysis of the inference performance metrics of each model replica within the most recent time window; The weight parameters of the lightweight neural network model are updated by calculating a weighted average based on the inference performance metrics of each model replica.

8. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, The edge sub-model corresponds to the shallow feature extraction part of the lightweight neural network model, the side sub-model corresponds to the middle feature extraction part of the lightweight neural network model, and the cloud sub-model corresponds to the deep feature extraction part and the task inference part of the lightweight neural network model. The terminal-side sub-model, the edge-side sub-model, and the cloud-side sub-model complete data interaction through intermediate feature data; the amount of intermediate feature data is less than the amount of original data in the real-time business data stream, and the intermediate feature data cannot be reversed to recover the original data of the real-time business data stream, thus ensuring data privacy and security.

9. The real-time analysis and processing method for business data based on lightweight neural networks according to claim 1, characterized in that, The S2 data preprocessing step also includes redundant data optimization processing: For video data in the real-time business data stream, motion gating frame selection technology is used to filter redundant frames, and only the frame data whose picture content changes is sent to the real-time inference step. Based on the data change rate of the real-time business data stream, the data sampling frequency is dynamically adjusted using adaptive sampling frequency adjustment technology.

10. A real-time business data analysis and processing system based on a lightweight neural network, used to implement the real-time business data analysis and processing method based on a lightweight neural network as described in any one of claims 1 to 9, characterized in that, include, The data acquisition module is used to acquire real-time business data streams generated by the business system; The data preprocessing module is used to perform data cleaning, format unification, timestamp alignment, redundant frame filtering, and adaptive sampling processing on the real-time business data stream. The model compression module is used to perform lightweight compression of the initial neural network model through a combination of strategies including structured pruning, knowledge distillation, and mixed precision quantization. The model segmentation and deployment module is used to segment lightweight neural network models based on multi-dimensional constraints and to deploy sub-models on the edge, cloud, and device sides. The edge-side inference engine, deployed on the terminal device, is used to perform the first-level real-time inference and output the first inference result and the corresponding confidence score; The hierarchical scheduling module is used to schedule inference tasks at different levels based on the comparison results of the confidence scores of the inference results with preset thresholds. The edge inference engine, deployed on edge servers, is used to perform second-level inference tasks; The cloud-based inference engine, deployed on a cloud server, is used to perform Level 3 inference tasks and output the final analysis results. The online adaptive update module is used to detect the concept drift phenomenon in business data streams and complete the online incremental update of the lightweight neural network model based on real-time business data.

Images