An end-cloud cooperative intelligent sales board system and implementation method
Patent Information
- Application Number
- CN202610922804.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-25
- Publication Date
- 2026-09-15
Smart Images
Figure CN122760147A_ABST
Abstract
Description
Technical Field
[0002] This invention relates to the fields of business intelligence and data processing technology, and in particular to an intelligent sales dashboard system and implementation method with edge-cloud collaboration. Background Technology
[0003] As enterprises deepen their digital transformation, sales dashboards have become a key tool for monitoring operational status and supporting decision-making. However, existing sales dashboard systems face many challenges in practical applications.
[0004] First, sales data typically originates from multiple heterogeneous channels, including online stores, offline retail outlets, and customer relationship management systems. These data differ in format, time granularity, and update frequency, and more importantly, their quality varies significantly. Existing systems often employ simple data integration methods, failing to effectively differentiate and address issues such as missing data and noise, resulting in low-quality data severely interfering with the accuracy of subsequent analysis.
[0005] Secondly, traditional anomaly detection methods often rely on fixed statistical thresholds, making it difficult to adapt to the complexity of sales scenarios. For example, during promotional activities or holidays, sharp fluctuations in sales are normal but can easily be misjudged as abnormal; while for genuine anomalies such as a slow decline in sales caused by inventory shortages, traditional methods may not be sensitive enough, leading to frequent false alarms and missed alarms.
[0006] Furthermore, managers at different levels have different focuses and analytical granularities regarding sales data, while existing systems often provide a single forecasting model and fixed forecasting dimensions, making it difficult to simultaneously meet the needs of multi-level and multi-dimensional analysis.
[0007] Finally, most existing Kanban systems follow a fixed "backend computing, frontend display" model, failing to consider the performance differences of user terminals and fluctuations in the network environment. When accessing from mobile devices or with poor network conditions, loading complex Kanban boards can lead to response delays and interactive stutters, severely impacting the user experience. Therefore, existing technologies urgently need improvement. Summary of the Invention
[0008] The purpose of this invention is to provide an intelligent sales dashboard system and implementation method that integrates edge and cloud computing, in order to solve the problems of low data fusion quality, inaccurate anomaly detection, single prediction dimension, and poor user interaction experience in the existing technology.
[0009] To address the aforementioned technical issues, this invention provides an edge-cloud collaborative intelligent sales dashboard system, comprising: a multi-source heterogeneous data acquisition and alignment module, used to calculate data quality indicators based on at least one of the following: data missing rate, noise level, and correlation with the main sales channel of each data source, and to perform weighted aggregation of the data using the data quality indicators as weights; an intelligent analysis engine, used to decompose the sales time series into trend items, periodic items, and residual items using a seasonal decomposition algorithm, to predict the residual items using a fusion prediction model to obtain residual item prediction results, and to perform anomaly detection on the residual items to generate preliminary anomaly judgments, and then to correct the preliminary anomaly judgments in conjunction with business rules to obtain anomaly judgment results; a fusion prediction module, used to output at least two of the following prediction results—daily sales prediction, weekly sales prediction, and stockout risk prediction—based on the trend item, periodic item, and residual item prediction results, using a hybrid neural network architecture including convolutional layers and recurrent neural network layers; and an adaptive visualization interaction module, used to dynamically determine the boundary between front-end and back-end calculations based on at least one of the following: terminal device performance, network conditions, and user roles.
[0010] Optionally, the data quality index is calculated using the formula Qi=α×(1-Mi)+β×(1-Ni)+γ×Ci, where Qi is the data quality index, Mi is the data missing rate, Ni is the noise level, Ci is the correlation coefficient with the main sales channel, and α, β, and γ are preset weights.
[0011] Optionally, the intelligent analysis engine is specifically used to: construct a dynamic confidence interval based on the statistical distribution of the prediction error generated by the prediction of the residual term; and generate the preliminary anomaly determination when the actual residual value is detected to exceed the dynamic confidence interval.
[0012] Optionally, the fusion prediction model for predicting the residual term includes a linear prediction model and a nonlinear prediction model, the weights of which are dynamically determined based on their mean square error on the validation set.
[0013] Optionally, the business rules include at least one of promotional activity rules, holiday rules, and inventory change rules.
[0014] Optionally, the hybrid neural network architecture includes multi-scale one-dimensional convolutional layers and long short-term memory network layers.
[0015] Optionally, the multi-scale one-dimensional convolutional layer includes at least two types of convolutional layers with convolutional kernels of different scales.
[0016] Optionally, the adaptive visualization interaction module is further configured to: calculate the edge-cloud partitioning score based on at least one of the terminal's CPU utilization, memory usage, and network round-trip latency; and determine the partitioning boundary based on the edge-cloud partitioning score.
[0017] Optionally, it also includes an intelligent decision support module, which is used to generate tiered early warnings, including prompt markers, highlighted notifications, SMS notifications, or email notifications, based on the anomaly judgment results obtained by the intelligent analysis engine.
[0018] This invention also provides a method for implementing an intelligent sales dashboard with edge-cloud collaboration, comprising: calculating a data quality index based on at least one of the data missing rate, noise level, and correlation with the main sales channel of each data source, and weighting and aggregating the data using the data quality index as a weight; decomposing the sales time series into trend, periodic, and residual terms using a seasonal decomposition algorithm, predicting the residual terms using a fusion prediction model to obtain residual term prediction results, performing anomaly detection on the residual terms to generate a preliminary anomaly judgment, and then correcting the preliminary anomaly judgment in conjunction with business rules to obtain anomaly judgment results; outputting at least two of the following prediction results—daily sales forecast, weekly sales forecast, and stockout risk forecast—based on the prediction results of the trend, periodic, and residual terms, using a hybrid neural network architecture including convolutional layers and recurrent neural network layers; and dynamically determining the boundary between front-end and back-end computation based on at least one of terminal device performance, network conditions, and user roles.
[0019] Compared with existing technologies, this patent has the following advantages: The robustness of data processing is improved: by calculating data quality indicators based on data missing rate, noise level and correlation, and using these as weights for weighted aggregation, the interference of low-quality data sources on the overall analysis results can be effectively suppressed, making the final fused data results more reliable and stable.
[0020] Reduced false alarm rate for anomaly detection: By decomposing the sales sequence and analyzing the more stable residuals, and by correcting the judgment results in conjunction with business rules such as promotions and holidays, it is possible to effectively distinguish between normal business fluctuations and real sales anomalies, significantly reducing the situation where large fluctuations caused by business activities are misjudged as anomalies.
[0021] It meets multi-level prediction needs: By adopting a multi-task hybrid neural network architecture that includes convolutional layers and recurrent neural network layers, it can simultaneously output prediction results of different dimensions and granularities, such as daily, weekly, and stockout risk, in a unified model, which efficiently meets the differentiated analysis needs of different management roles in enterprises.
[0022] Improved Kanban responsiveness and user experience: By dynamically adjusting the boundaries of computing tasks between the front-end and back-end based on information such as terminal performance, network conditions, and user roles, the computing load can be intelligently distributed between the cloud and the terminal, ensuring a smooth and timely interactive experience under different device and network conditions, and avoiding the performance bottlenecks of traditional fixed architectures. Attached Figure Description
[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a schematic diagram of the structure of an intelligent sales dashboard system according to an embodiment of the present invention.
[0025] Figure 2 This is a schematic diagram of the workflow of the multi-source heterogeneous data acquisition and alignment module according to an embodiment of the present invention.
[0026] Figure 3 This is a schematic diagram illustrating the working principle of the intelligent analysis engine according to an embodiment of the present invention.
[0027] Figure 4 This is a schematic diagram of the fusion prediction module according to an embodiment of the present invention.
[0028] Figure 5 This is a schematic diagram of the adaptive visualization interaction module performing edge-cloud collaborative decision-making according to an embodiment of the present invention.
[0029] Figure 6 This is a flowchart illustrating the implementation method of an intelligent sales dashboard according to an embodiment of the present invention.
[0030] Figure 7 This is a timing diagram illustrating edge-cloud collaborative interaction in a weak network environment according to an embodiment of the present invention.
[0031] The reference numerals in the attached figures are explained as follows: 100: Intelligent Sales Dashboard System; 110: Multi-source Heterogeneous Data Acquisition and Alignment Module; 120: Data Processing and Feature Engineering Module; 130: Intelligent Analysis Engine; 140: Fusion Prediction Module; 150: Adaptive Visualization Interaction Module; 160: Intelligent Decision Support Module; 240: Quality Indicator Calculation Unit; 250: Weighted Aggregation Unit; 260: Fusion Data; 301: Sales Time Series; 310: Seasonal Decomposition; 320: Residual Term; 330: Fusion Prediction Model; 340: Dynamic Confidence Interval Construction; 350: Business Rule Base; 360: Preliminary Anomaly Detection; 370: Anomaly Detection Result; 401: Fusion Time Series Data; 420: Multi-scale One-dimensional Convolutional Layer; 430: Long Short-term Memory (LSTM) Layer; 440: Daily sales forecast header; 450: Weekly sales forecast header; 460: Stockout risk forecast header; 510: Client terminal; 520: Backend server; 530: Terminal performance monitoring; 540: Network status monitoring; 550: End-to-cloud collaborative decision-making unit; 560: Computing task allocation; 701: User terminal; 702: Frontend module; 703: Backend server. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by those skilled in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application. Before detailing the embodiments of this application, the nouns and terms involved in the embodiments are explained as follows: (1) Data quality index: refers to a comprehensive index used to quantitatively evaluate the reliability of data sources. Its calculation basis includes at least one of the following: data missing rate, noise level, and correlation with the main sales channel. In this invention, this index is used as a weight in the weighted aggregation process to ensure that high-quality data plays a dominant role in the analysis, thereby improving the accuracy and robustness of subsequent analysis and prediction.
[0033] (2) Preliminary anomaly judgment: This refers to the original anomaly judgment result generated by the intelligent analysis engine solely through the algorithm model (e.g., by comparing the actual residual value with the dynamically constructed confidence interval) and has not yet been corrected by business rules. This result is an intermediate step in distinguishing between purely statistical fluctuations and fluctuations with specific business backgrounds, providing a foundation for subsequent accurate judgments based on business knowledge.
[0034] (3) Boundary Delineation: In an edge-cloud collaborative computing architecture, this refers to the dynamic decision boundary used to define which computing tasks are executed on the client (front end) and which are executed on the server (back end). The determination of this boundary is based on at least one of the following: terminal device performance, network conditions, and user role. Its purpose is to optimize the overall allocation of computing resources of the system while ensuring a smooth user experience.
[0035] Please see Figure 1 This application provides an edge-cloud collaborative intelligent sales dashboard system 100, aiming to solve technical problems in the prior art such as inconsistent quality of sales data due to diverse sources, susceptibility of anomaly detection to business fluctuations, single prediction dimension, and poor user experience across different terminals. This system achieves end-to-end optimization from data access to intelligent analysis and adaptive presentation through a modular design that facilitates collaborative operation.
[0036] The system provided in this application embodiment is based on a multi-source heterogeneous data acquisition and alignment module 110, which enables quality-aware fusion of data from different sources. Specifically, this module does not simply aggregate data; instead, it first calculates a quantitative data quality index based on at least one of the following: data missing rate, noise level, and correlation with the main sales channel for each data source. This design aims to abandon the crude approach of treating all data sources equally in traditional technologies, providing a scientific weighting basis for subsequent aggregation. Subsequently, the module uses the calculated data quality index as weights to perform weighted aggregation of the data. In this way, the system can automatically amplify the influence of high-quality data sources while suppressing interference from low-quality data sources, fundamentally solving the problem of inaccurate analysis results caused by data quality issues and significantly improving the overall reliability of the input data.
[0037] After obtaining high-quality fused data, the system performs in-depth analysis using an intelligent analysis engine 130. This engine first uses seasonal decomposition 310 to break down the complex sales time series 301 into more easily analyzable trend, cyclical, and residual terms 320. Existing technologies typically perform anomaly detection directly on the original series, easily misjudging normal seasonal fluctuations or growth trends as anomalies. The purpose of this solution is to focus on analyzing the irregular residual terms 320 by stripping away predictable trend and cyclical components, thereby more accurately capturing true anomaly signals. Next, the engine uses a fusion prediction model 330 to predict the residual terms 320, obtaining the predicted residual terms, and then performs anomaly detection on the actual residual terms 320 to generate a preliminary anomaly determination 360. Subsequently, the system does not stop at purely algorithmic judgment but further refines the preliminary anomaly determination 360 by incorporating rules from the business rule base 350, ultimately obtaining the anomaly determination result 370. For example, a surge in sales identified by an algorithm might be corrected to a normal fluctuation if, after combining it with business rules, it is found to align with the "Double Eleven" promotional event. This combination of algorithm and business knowledge significantly reduces the false alarm rate of anomaly detection, enabling precise insights into sales dynamics.
[0038] To meet the decision-making needs of different management levels within an enterprise, the system also includes a fusion prediction module 140. This module is designed to overcome the limitations of traditional single prediction models that cannot simultaneously handle different time granularities and analytical dimensions. Specifically, based on the prediction results of the decomposed trend term, period term, and the aforementioned residual term, this module processes the data through a hybrid neural network architecture containing convolutional layers and recurrent neural network layers. This hybrid architecture can simultaneously capture both short-term local patterns (handled by convolutional layers) and long-term dependencies (handled by recurrent neural network layers) in sales data. Ultimately, this module can output at least two prediction results in parallel from daily sales forecasts, weekly sales forecasts, and stockout risk forecasts. Through this multi-task learning approach, a single system can simultaneously serve store managers focused on daily operations and regional directors focused on weekly planning, achieving high efficiency and reusability of the prediction function.
[0039] Finally, to address the issue of poor user experience across diverse terminal devices and network environments, the system introduces an adaptive visual interaction module 150. The core design principle of this module is to decouple front-end display from back-end computation and establish a dynamic collaboration mechanism. The module continuously monitors terminal device performance (e.g., CPU and memory usage), network conditions (e.g., latency and bandwidth), and the currently logged-in user's role. Based on this contextual information, the module dynamically determines the boundary between front-end and back-end computation. For example, on a PC with a good network, lightweight computational tasks such as data aggregation and filtering can be performed on the front-end to accelerate interactive response; while on mobile devices with poor network conditions, most computational tasks are performed on the back-end server 520, with the front-end only responsible for rendering the final result. This dynamic adjustment of computational load allocation solves the problems of lag and latency in traditional fixed architectures under weak conditions, ensuring a smooth and efficient dashboard interaction experience for users in any scenario.
[0040] Furthermore, in a preferred embodiment, the multi-source heterogeneous data acquisition and alignment module 110 employs a more specific quantification method when calculating data quality indicators. Please refer to [link / reference]. Figure 2 The quality indicator calculation unit 240 within the module can calculate the data quality indicator Qi according to the following formula: Qi = α × (1 - Mi) + β × (1 - Ni) + γ × Ci; Where Qi is the data quality index of the i-th data source, Mi is the data missing rate of the data source within the statistical period, Ni is its noise level (e.g., measured by the deviation of a data point from the mean of its neighboring points), Ci is the correlation coefficient between the sales sequence of the data source and the sales sequence of the main sales channel (e.g., the company's total sales), and α, β, and γ are preset weighting coefficients, satisfying α+β+γ=1. This formulaic definition makes the data quality assessment process objective and reproducible. By adjusting the weighting coefficients α, β, and γ, enterprises can, according to their own business characteristics, focus on penalizing channels with high missing rates or place more emphasis on consistency with the overall trend. This specific implementation method makes data quality assessment no longer a vague concept, but a configurable and optimizable engineering step. Its technical effect is to provide more accurate and controllable weight inputs for the subsequent weighted aggregation unit 250, thereby further improving the quality of the fused data 260.
[0041] It should be noted that, in a preferred embodiment, the anomaly detection logic of the intelligent analysis engine 130 has been refined. Please refer to [link / reference]. Figure 3After predicting the residual term 320 using the fusion prediction model 330, the engine does not use a fixed threshold for judgment. Instead, it dynamically constructs a confidence interval based on the statistical distribution of the prediction error generated by the prediction model on historical data (e.g., calculating the standard deviation of the error). This step is completed by the dynamic confidence interval construction unit 340. The width of this dynamic confidence interval adaptively adjusts with recent data fluctuations. For example, during periods of stable data, the interval is narrower for more sensitive detection; during periods of data fluctuation, the interval is appropriately widened to tolerate larger normal fluctuations. Subsequently, when the engine detects that the actual residual value at a certain point in time exceeds the range of this dynamic confidence interval, it generates a preliminary anomaly judgment 360. This method based on dynamic confidence intervals, compared to the traditional fixed threshold method, can better adapt to changes in the data's own patterns. Its technical effect is to significantly improve the accuracy and adaptability of the preliminary anomaly judgment, providing higher-quality input for subsequent business rule corrections.
[0042] In another preferred embodiment, the internal structure of the aforementioned fusion prediction model 330 for predicting residual terms is specifically defined. This model may include a linear prediction model (such as an ARIMA model) and a nonlinear prediction model (such as a gradient boosting tree or neural network model). These two models have different design focuses: linear models excel at capturing regular autocorrelation patterns in data, while nonlinear models are better at fitting complex, unstructured fluctuation patterns. To combine their advantages, the system does not simply add their predictions together, but dynamically determines their weights based on their respective performance on the historical validation set. Specifically, the weights of the linear and nonlinear models can be dynamically determined based on their mean squared error (MSE) on the validation set. This dynamic weighting strategy allows the model to automatically favor the model that performs better in the current scenario, depending on changes in data characteristics. For example, when the residual sequence exhibits a clear linear pattern, the weight of the linear model will naturally increase. In this way, the overall prediction accuracy and generalization ability of the fusion prediction model are effectively improved.
[0043] In one optional implementation, the contents of the business rule base 350 upon which the intelligent analysis engine 130 relies are exemplarily described. The business rules may include at least one of promotional activity rules, holiday rules, and inventory change rules. Promotional activity rules can define the time windows for all known marketing activities of the company, such as "618" and "Double Eleven." Holiday rules contain information on statutory holidays and industry-specific peak sales seasons. Inventory change rules record important replenishment, clearance, or stockout events. When the preliminary anomaly judgment 360 generated by the intelligent analysis engine 130 coincides with the time point of these business rules, the system executes correction logic. For example, if a preliminary anomaly judgment of a sales peak occurs during "Double Eleven," the system will correct it to normal business fluctuations. By introducing these specific business rules, the system's anomaly detection is no longer a purely mathematical game but truly integrated with business logic. Its technical effect lies in its ability to accurately distinguish between statistical anomalies and business normalities, thereby providing managers with truly valuable anomaly signals.
[0044] Furthermore, in a preferred embodiment, the hybrid neural network architecture in the fusion prediction module 140 is described in more detail. Please refer to [link / reference]. Figure 4 The architecture can specifically include a multi-scale one-dimensional convolutional layer 420 and a Long Short-Term Memory (LSTM) layer 430. The multi-scale one-dimensional convolutional layer 420 is designed to extract local features across different time spans from the fused time-series data 401. For example, using a small convolutional kernel (e.g., a window size of 3) can capture intraday sales fluctuation patterns, while using a large convolutional kernel (e.g., a window size of 14) can capture weekly sales rhythms. These features extracted from different scales are then fed into the LSTM layer 430. LSTM layers, as a special type of recurrent neural network, excel at processing and predicting events with very long intervals and delays in time series. They are able to learn long-term dependencies between local patterns extracted by the convolutional layers. By combining the local feature extraction capability of CNN with the long-term temporal modeling capability of LSTM, this hybrid neural network architecture can more comprehensively understand the inherent patterns of sales data, thereby providing high-quality feature representations for subsequent multi-task predictions (such as daily sales forecast head 440, weekly sales forecast head 450, and stockout risk prediction head 460), significantly improving the accuracy of predictions.
[0045] In another preferred embodiment, the configuration of the aforementioned multi-scale one-dimensional convolutional layer 420 is further defined. This layer specifically includes at least two convolutional layers with kernels of different scales. For example, a convolutional layer with a kernel size of 3, a convolutional layer with a kernel size of 7, and a convolutional layer with a kernel size of 14 can be configured in parallel. These three convolutional layers operate on the input time series, aiming to capture sales patterns within time windows of 3 days, 7 days (one week), and 14 days (two weeks), respectively. The principle behind this multi-scale design is that effective information in sales data may exist at different time scales, and a single-scale convolutional kernel is prone to overlooking some aspects. By extracting multi-scale features in parallel and then fusing them, the model can act like an observer with multiple pairs of glasses with different focal lengths, simultaneously seeing close-up details and distant outlines, thereby more completely capturing the key factors driving sales changes. The technical effect is to further enhance the model's ability to represent complex sales patterns.
[0046] In one alternative implementation, the decision-making process by which the adaptive visualization interaction module 150 determines the boundary division is detailed. See also... Figure 5 This module can incorporate an edge-cloud collaborative decision-making unit 550. This unit first calculates a quantified "edge-cloud partitioning score" based on real-time data obtained from the terminal performance monitoring module 530 and network status monitoring module 540 of the client terminal 510. This score calculation comprehensively considers the terminal's CPU utilization, memory usage, and network round-trip time (RTT) with the backend server 520. For example, lower CPU and memory usage, and lower network latency result in a higher score. Subsequently, the decision-making unit 550 determines the partitioning boundary based on the edge-cloud partitioning score and allocates computing tasks 560 accordingly. For example, a score threshold can be preset. When the score is higher than this threshold, the system determines that the current terminal environment is good and allocates more computing tasks (such as data filtering and sorting) to the front end; otherwise, these tasks are retained for execution on the backend server 520. By introducing the quantified indicator of "edge-cloud partitioning score," the edge-cloud collaborative decision-making process is no longer a vague judgment but a data-driven, dynamically adjustable closed-loop control system, thereby achieving precise management of computing load allocation.
[0047] It should be noted that the system in this embodiment may also include an intelligent decision support module 160. This module is connected to the intelligent analysis engine 130 and is designed to transform the analysis results into actionable actions or clear warnings. Specifically, this module generates tiered warnings based on the anomaly judgment result 370 finally output by the intelligent analysis engine 130. For example, for anomalies that have just occurred and have little impact, the system may only issue a level one warning by displaying a prompt mark or highlighting the anomaly on the dashboard; for anomalies that are continuously occurring and may pose risks, the system will issue a level two warning through in-system messages or highlighted notifications; and for anomalies that are judged to be serious and have a worsening trend, the system will trigger a level three warning and notify the relevant personnel via SMS or email. This tiered warning mechanism avoids information overload and helps managers quickly focus on the most important issues. Its technical effect is to bridge the "last mile" from data analysis to decision response, making the entire system form a closed loop of "discovery-analysis-warning".
[0048] Please see Figure 6 This application also provides a method for implementing an intelligent sales dashboard with end-to-end cloud collaboration. This method can run in the system described in any of the above embodiments. The method begins at step 601 and specifically includes the following steps: Step S602: Perform data collection and aggregation. The system collects sales-related data from multiple different data sources (such as ERP and CRM systems) through the multi-source heterogeneous data collection and alignment module 110. In this step, the system calculates data quality indicators based on dimensions such as data missing rate, noise level, and relevance to the main sales channel of each data source, and uses these indicators as weights to perform weighted aggregation of the data, forming a unified, high-quality sales time series.
[0049] Step S603: Perform data processing and feature engineering. The system's data processing and feature engineering module 120 will process the aggregated data, for example, to construct features containing multi-dimensional information such as time, promotion, and external environment, in preparation for subsequent analysis and prediction.
[0050] Step S604: Perform anomaly detection and correction. The intelligent analysis engine 130 receives the processed data and uses a seasonal decomposition algorithm to decompose the sales time series into trend items, periodic items, and residual items. Next, it uses a fusion prediction model to predict the residual items and performs anomaly detection on the actual residual items, generating a preliminary anomaly judgment. Finally, it refines the preliminary judgment by incorporating business rule bases (such as promotion and holiday rules) to obtain the final anomaly judgment result.
[0051] Step S605: Perform multi-task prediction. The fusion prediction module 140 utilizes its internal hybrid neural network architecture (such as CNN-LSTM) to output multiple prediction results in parallel, including daily sales forecast, weekly sales forecast, and stockout risk prediction, based on the processed feature data.
[0052] Step S606: Perform edge-cloud collaborative decision-making. The adaptive visualization interaction module 150 dynamically determines the boundary between front-end computing and back-end computing based on real-time monitored user terminal performance, network conditions, and user roles, in order to optimize the user experience.
[0053] Step S607: Perform visualization and early warning. The system will visualize the anomaly detection results and multi-task prediction results on the front end. At the same time, the intelligent decision support module 160 will generate and push corresponding level of early warning information (such as prompt markers, highlighted notifications, SMS or email) according to the severity of the anomaly.
[0054] It should be noted that the process ends at step 608. By executing the above method, problems such as data quality, false alarms, limited predictions, and poor user experience in existing technologies can be systematically resolved.
[0055] To more vividly illustrate the technical effects of this invention, a specific application scenario will be described below. Suppose a regional sales manager is on a high-speed train with poor signal and needs to access the company's smart sales dashboard via their mobile phone to stay updated on the latest sales developments. Please refer to... Figure 7 This is a schematic diagram of the interaction sequence in this scenario.
[0056] First, the user initiates a data view request through a browser on their terminal 701. This request is received by the front-end module 702 running in the browser. On the one hand, the front-end module 702 requests basic data and computing resources from the back-end server 703; on the other hand, its built-in monitoring logic starts working, continuously monitoring the current network status and terminal performance.
[0057] During the interaction, the front-end module 702 detected a network round-trip time (RTT) as high as 800 milliseconds, far exceeding the normal level. Based on this information, the end-cloud collaborative decision-making unit 550 in the adaptive visualization interaction module 150 determined that the current network environment was weak and dynamically switched the working mode to "thin client" mode. In this mode, the front-end module 702 no longer requests the raw aggregated data that requires a lot of computation from the back-end server 703, but instead sends a clear instruction to request that all heavy computational tasks, such as chart rendering and secondary data aggregation, be completed on the back-end.
[0058] Upon receiving the instruction, backend server 703 mobilizes its powerful computing resources to execute all computational tasks, even directly rendering the complex charts to be displayed into a static image or a piece of minimal HTML code on the server side. Then, backend server 703 returns this lightweight final rendering result to frontend module 702.
[0059] Finally, after receiving the image or HTML fragment, the front-end module 702 displays it directly on the screen of the user terminal 701 without performing complex calculations and rendering. In this way, even in scenarios with extremely poor network conditions, users can see the most crucial sales data charts in the shortest possible time, instead of facing a long blank screen or a laggy interactive interface. This embodiment fully demonstrates the technical advantages of this invention, which dynamically adjusts the boundaries of edge cloud computing to ensure the availability of core functions and a good user experience even in harsh environments.
[0060] In another embodiment of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, is capable of performing the above-described functions. Figure 6 The method illustrates all or part of the steps in the edge-cloud collaborative intelligent sales dashboard implementation method. The storage medium can be any type of non-volatile memory, such as a solid-state drive (SSD), hard disk drive (HDD), optical disc, or flash drive. Furthermore, the invention also covers a computer device including a processor and the aforementioned computer-readable storage medium, configured to perform the methods provided by the invention.
[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An intelligent sales dashboard system with edge-cloud collaboration, characterized in that, include: The multi-source heterogeneous data acquisition and alignment module is used to calculate data quality indicators based on at least one of the following: data missing rate, noise level, and correlation with the main sales channel of each data source, and to perform weighted aggregation of the data using the data quality indicators as weights. The intelligent analysis engine is used to decompose the sales time series into trend items, periodic items and residual items using a seasonal decomposition algorithm, use a fusion prediction model to predict the residual items to obtain the residual item prediction results, and perform anomaly detection on the residual items to generate a preliminary anomaly judgment. Then, the preliminary anomaly judgment is corrected by combining business rules to obtain the anomaly judgment result. The fusion prediction module is used to output at least two of the following prediction results—daily sales forecast, weekly sales forecast, and stockout risk forecast—based on the prediction results of the trend term, the periodic term, and the residual term, through a hybrid neural network architecture including convolutional layers and recurrent neural network layers. The adaptive visualization interaction module is used to dynamically determine the boundary between front-end computing and back-end computing based on at least one of the following: terminal device performance, network conditions, and user roles.
2. The system according to claim 1, characterized in that, The data quality index is calculated using the formula Qi=α×(1-Mi)+β×(1-Ni)+γ×Ci, where Qi is the data quality index, Mi is the data missing rate, Ni is the noise level, Ci is the correlation coefficient with the main sales channel, and α, β, and γ are preset weights.
3. The system according to claim 1, characterized in that, The intelligent analysis engine is specifically used for: A dynamic confidence interval is constructed based on the statistical distribution of the prediction error generated by the prediction of the residual term; and When the actual residual value is detected to exceed the dynamic confidence interval, the preliminary anomaly determination is generated.
4. The system according to claim 3, characterized in that, The fusion prediction model used to predict the residual term includes a linear prediction model and a nonlinear prediction model, the weights of which are dynamically determined based on their mean square error on the validation set.
5. The system according to claim 1, characterized in that, The business rules include at least one of the following: promotional activity rules, holiday rules, and inventory change rules.
6. The system according to claim 1, characterized in that, The hybrid neural network architecture includes multi-scale one-dimensional convolutional layers and long short-term memory network layers.
7. The system according to claim 6, characterized in that, The multi-scale one-dimensional convolutional layer includes at least two types of convolutional layers with convolutional kernels of different scales.
8. The system according to claim 1, characterized in that, The adaptive visualization interaction module is also used for: The computing cloud is scored based on at least one of the following: CPU utilization, memory usage, and network round-trip latency. The division boundary is determined based on the edge-cloud division score.
9. The system according to claim 1, characterized in that, It also includes an intelligent decision support module, which is used to generate tiered early warnings, including prompts, highlighted notifications, SMS notifications or email notifications, based on the anomaly judgment results obtained by the intelligent analysis engine.
10. A method for implementing an intelligent sales dashboard with edge-cloud collaboration, characterized in that, include: The steps are as follows: Calculate a data quality index based on at least one of the following: data missing rate, noise level, and correlation with the main sales channel; and use the data quality index as a weight to perform weighted aggregation of the data. The steps are as follows: using a seasonal decomposition algorithm to decompose the sales time series into trend items, periodic items and residual items; using a fusion prediction model to predict the residual items to obtain the residual item prediction results; performing anomaly detection on the residual items to generate a preliminary anomaly judgment; and then combining business rules to correct the preliminary anomaly judgment to obtain the anomaly judgment result. Based on the prediction results of the trend term, the periodic term, and the residual term, the step of outputting at least two of the prediction results of daily sales forecast, weekly sales forecast, and stockout risk forecast through a hybrid neural network architecture including convolutional layers and recurrent neural network layers; The steps for dynamically determining the boundary between front-end computing and back-end computing based on at least one of the following: terminal device performance, network conditions, and user roles.