A management system supporting real-time aggregation and intelligent classification of multi-store orders

Abstract

The application belongs to the technical field of computer data processing, and relates to a management system supporting real-time convergence and intelligent classification of multi-store orders, which comprises: a monitoring and strategy generation module that acquires global operation state parameters and generates adaptive coding strategy instructions according to a business scene model; an edge collection module that captures original order messages and extracts a target key field set according to the instructions; an edge coding module that codes the field set to generate a variable-structure order feature vector; a receiving and loading module that receives the vector and activates a matching target decoding head model; a decoding restoration module that generates a standardized decision feature tensor by using the model; a classification calculation module that outputs intelligent classification result labels by using a multi-level classification network; and a feedback and updating module that executes business control logic and updates parameters. The application solves the problem of dynamically adjusting edge data processing strategies according to real-time system states in the multi-store order processing scene to balance data processing accuracy and system resource consumption.

Description

Technical Field

[0001] This invention belongs to the technical field of computer data processing and relates to a management system that supports real-time aggregation and intelligent classification of orders from multiple stores. Background Technology

[0002] In the e-commerce field, businesses typically connect to multiple e-commerce platforms to cover a wide range of sales channels. Because different e-commerce platforms have independent order data formats and interface specifications, order data from different stores exhibits significant structural differences. To achieve unified management of multi-source heterogeneous data, traditional order processing systems usually employ two mainstream data processing architectures: one is to uniformly receive, parse, and clean all raw order data on a central server; the other is to pre-deploy fixed data extraction rules at the edge, closer to the data source, to move some processing tasks forward.

[0003] Existing centralized processing solutions on the central server side rely on transmitting all raw order messages to a central node in real time for subsequent classification and decision-making, while edge preprocessing solutions rely on pre-defined static field extraction logic and fixed encoding rules to filter data. These two solutions focus on high-performance centralized processing at the back end or fixed-rule filtering at the front end, respectively, attempting to address the aggregation needs of multi-store orders by increasing transmission bandwidth or reducing data transmission volume.

[0004] However, the aforementioned existing technologies have limitations in dealing with dynamically changing business scenarios. Centralized processing solutions, in high-concurrency scenarios such as large-scale promotional events, can strain network bandwidth due to the centralized transmission of massive amounts of raw data, and the computing and storage resources of the central server are prone to bottlenecks, leading to reduced order fulfillment timeliness. Edge-end fixed-rule solutions, on the other hand, struggle to adapt to real-time fluctuations in system load and business status. Under low system load, rule limitations may prevent the transmission of sufficient feature data to support refined risk control, while under high load, they cannot automatically reduce data dimensions to ensure core processes, resulting in data transmission redundancy or missing key features. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a management system that supports real-time aggregation and intelligent classification of orders from multiple stores.

[0006] A management system that supports real-time aggregation and intelligent categorization of orders from multiple stores, comprising: The monitoring and strategy generation module is used to obtain global operational status parameters of the multi-channel e-commerce system and generate adaptive coding strategy instructions based on the preset business scenario model. The edge acquisition module uses edge acquisition probes to capture the original order message and extracts the target key field set according to the adaptive encoding strategy instructions; The edge encoding module, based on the vector construction logic defined by the adaptive encoding strategy instruction, performs structured encoding on the target key field set to generate a variant order feature vector; The receiving and loading module uses the central processing unit to receive the variant order feature vector and activates the target decoding head model that matches the current strategy; The decoding and restoration module uses the target decoding head model to perform semantic restoration and spatial projection on the variant order feature vector, generating a standardized decision feature tensor. The classification calculation module inputs the decision feature tensor into a multi-level classification network, calculates the business attributes and risk level of the order, and outputs intelligent classification result labels. The feedback and update module responds to the intelligent classification result labels, executes business control logic, and updates global operational status parameters to form a closed-loop control.

[0007] A further aspect of the present invention includes a monitoring and policy generation module, which performs the following steps: By monitoring the system load rate, the promotional activity status of specific products, and the flow control threshold of the platform interface through the central control node, global operational status parameters are generated. Input the global operational status parameters into the preset strategy generator to calculate the overall system status score; When the overall system status score exceeds the preset threshold, an adaptive encoding strategy instruction containing field extraction rules and vector space definition is generated and sent to the edge acquisition probe.

[0008] A further aspect of the present invention includes an edge acquisition module, used to perform the following operations: Use edge acquisition probes to monitor order generation events on e-commerce platforms in real time and capture raw order messages containing full metadata; Parse the field extraction rules in the adaptive encoding strategy instruction and apply the path expression to the original order message; Filter out unnecessary information and select numerical values ​​or strings that match the current business scenario to form a target key field set with a variable length as the strategy changes.

[0009] A further aspect of the present invention includes an edge coding module, which performs the following steps: Iterate through the target key field set, perform hash mapping processing on discrete data, and perform normalization processing on continuous numerical data; The processed data is assembled into a one-dimensional array according to the dimensional arrangement order defined by the adaptive encoding strategy instructions; A version identifier corresponding to the adaptive encoding strategy instruction is inserted into the header of a one-dimensional array, and the modified order feature vector is generated by encapsulation.

[0010] A further embodiment of the present invention includes a receiving and loading module, configured to perform the following steps: The central processing unit parses the version identifier bit in the header of the modified order feature vector; The model file path corresponding to the version identifier is indexed in the preset model configuration registry; Load and instantiate the target decoder head model from the pre-trained multimodal decoder library. The target decoder head model is configured with a feature mapping matrix that is adapted to the current vector structure.

[0011] A further aspect of the present invention includes a decoding and restoration module, which performs the following operations: The modified order feature vector is input into the target decoding head model, and the hybrid feature mapping layer is used to perform table embedding for discrete dimensions and linear projection for continuous dimensions. The projection results are stitched together and restored to intermediate features with business semantics; The intermediate features are projected onto a unified high-dimensional feature space, and the dimensions not collected by the current strategy are filled with default values ​​to generate a decision feature tensor.

[0012] A further aspect of the present invention includes a classification calculation module, which performs the following operations: Input the decision feature tensor into the primary fast response classifier to calculate the basic priority probability of the order; Determine whether the basic priority probability falls within a preset fuzzy range; If the value falls within the fuzzy range, the deep neural network model is triggered to perform correlation analysis on the decision feature tensor, and the output values ​​of each level of the model are combined to generate intelligent classification result labels.

[0013] A further aspect of the present invention includes a feedback and update module, used to perform the following operations: Analyze the labels of intelligent classification results to identify risk types and processing priorities; Perform inventory pre-locking on orders marked as first priority; For orders marked with a specific risk type, execute an interception and suspension operation; For orders marked as requiring in-depth verification, initiate an asynchronous full data pull-back thread to obtain complete order details from the source.

[0014] A further aspect of this invention, updating the global operational status parameters specifically includes the following sub-steps: statistically analyzing the processing time and system resource consumption data during the order processing process; The current statistical data and historical data are weighted and merged using a smoothing coefficient to update the global operational status parameters; The updated global operational status parameters are compared with the preset scene switching threshold. If the threshold is exceeded, a new adaptive coding strategy instruction is generated.

[0015] A further aspect of the present invention includes the following specific methods for monitoring each parameter: The utilization rate of the central processing unit is obtained by reading the server performance counters as the system load rate; The current enumeration value of the product SKU is obtained by performing a database query operation as the status of the promotional activity; The remaining call count in the response header returned by the e-commerce platform API is used as the flow control threshold for the platform interface.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention monitors global operational parameters such as system load and promotional status, and uses a preset model to generate adaptive coding strategy instructions, which are then distributed to the edge. This mechanism enables the system to adjust the dimensions and granularity of data collection based on real-time operating conditions: automatically simplifying collected fields under high load scenarios to ensure smooth core business operations, and increasing collected dimensions under low load scenarios to support in-depth analysis. This dynamic adjustment mechanism balances the relationship between data analysis accuracy and system resource consumption, optimizing resource allocation efficiency under different business scenarios.

[0017] 2. This invention utilizes acquisition probes at the edge to perform source filtering and extraction of original order messages according to policy instructions. By removing redundant information irrelevant to the current decision before data transmission, the total amount of data transmitted to the central processing unit is reduced, thereby alleviating network bandwidth pressure. Simultaneously, by pre-positioning some data cleaning and filtering work at the edge, the computational overhead of the central server performing preliminary data cleaning after receiving the data is reduced.

[0018] 3. This invention maps source data with diverse structures and encoding methods into standardized decision feature tensors by combining modified order feature vectors with the target decoding head model. This process shields downstream processing logic from the impact of upstream multi-store data format differences and changes in collection strategies, providing a unified data input interface for the classification calculation module. This eliminates the need for separate adaptation of the classification model for each data source format or strategy change, facilitating model reuse and subsequent system expansion and maintenance.

[0019] 4. This invention establishes a feedback closed-loop mechanism based on business execution results, feeding back order processing timeliness, resource consumption, and risk identification results to the strategy generation end to update global operational status parameters. When system status changes reach preset conditions, the generation and issuance of new strategies are automatically triggered. This closed-loop control enables the system to continuously adjust data processing strategies based on actual operational results, improving the system's robustness and stability in the complex and ever-changing e-commerce business environment. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0021] Figure 1 This is a schematic diagram of the framework in the embodiments of this application.

[0022] Figure 2 This is a flowchart illustrating an embodiment of this application. Detailed Implementation

[0023] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0024] See attached document Figure 1 - Figure 2 This invention proposes a management system that supports real-time aggregation and intelligent classification of orders from multiple stores, including the following modules: The monitoring and strategy generation module is used to obtain global operational status parameters of the multi-channel e-commerce system and generate adaptive coding strategy instructions based on the preset business scenario model. The edge acquisition module uses edge acquisition probes to capture the original order message and extracts the target key field set according to the adaptive encoding strategy instructions; The edge encoding module, based on the vector construction logic defined by the adaptive encoding strategy instruction, performs structured encoding on the target key field set to generate a variant order feature vector; The receiving and loading module uses the central processing unit to receive the variant order feature vector and activates the target decoding head model that matches the current strategy; The decoding and restoration module uses the target decoding head model to perform semantic restoration and spatial projection on the variant order feature vector, generating a standardized decision feature tensor. The classification calculation module inputs the decision feature tensor into a multi-level classification network, calculates the business attributes and risk level of the order, and outputs intelligent classification result labels. The feedback and update module responds to the intelligent classification result labels, executes business control logic, and updates global operational status parameters to form a closed-loop control.

[0025] In one embodiment of the present invention, the monitoring and policy generation module is used to perform the following steps: By monitoring the system load rate, the promotional activity status of specific products, and the flow control threshold of the platform interface through the central control node, global operational status parameters are generated. Input the global operational status parameters into the preset strategy generator to calculate the overall system status score; When the overall system status score exceeds the preset threshold, an adaptive encoding strategy instruction containing field extraction rules and vector space definition is generated and sent to the edge acquisition probe.

[0026] Specifically, the monitoring module built into the central control node first concurrently acquires global operational status parameters of the multi-channel e-commerce system through the internal monitoring bus and external API interface at a preset polling cycle. It should be understood that these global operational status parameters are a structured dataset describing the overall operational status of the entire e-commerce system. The monitoring module obtains the utilization rate of the central processing unit as the system load rate by reading server performance counters. Its value range is typically from 0 to 100%; by executing an SQL query, the current promotional status of the product SKU can be obtained from the marketing campaign configuration database. This status is typically an enumerated value, such as 0 representing regular sales and 1 representing flash sales or limited-time offers; and the flow control threshold of the platform interface is obtained by parsing the X-RateLimit-Remaining field in the HTTP response header returned by the e-commerce platform API. This is the difference between the maximum number of calls allowed within a unit time period and the current remaining number of calls.

[0027] These heterogeneous raw monitoring data are aggregated and formatted into structured global operational status parameter data objects. Subsequently, the global operational status parameters are used as input and pushed in real-time to a pre-defined business scenario model. In this embodiment, the pre-defined business scenario model is a set of predefined rules used to map different combinations of global operational status parameters to specific coding strategies. The strategy generator calculates a comprehensive system status score based on the input indicator values. This is used to quantify whether to switch the coding strategy. The specific calculation logic is shown in the following formula:

[0028] in, This is the normalized system load rate, which is the current CPU utilization rate divided by 100%. This serves as a status indicator for promotional activities; it is set to 1 during high-concurrency activities and 0 otherwise. For interface flow control proximity, the calculation method is as follows: . , and These are the weighting coefficients for each state parameter, and their sum is 1. These three weighting coefficients are not arbitrarily chosen, but rather derived from historical traffic stress test data and fitted using a logistic regression model to determine their contribution to system crash risk. In this embodiment, for a high-concurrency online retail system with high stability requirements, the weights are set to... , and .

[0029] This model is essentially a policy generator based on the Drools rule engine. When the system state is comprehensively scored... When the preset scene switching threshold is exceeded, such as 0.85 based on historical stress test data statistics, a corresponding rule will be triggered, generating a lightweight and timely adaptive encoding strategy instruction. The adaptive encoding strategy instruction is a configuration file that guides the edge acquisition probes in data processing. It contains field extraction rules and vector space definitions to guide edge data processing. Field extraction rules are a set of path expressions used to locate and extract target data fields from the original order message; while the vector space definition details the dimensional composition of the modified order feature vector, the data type of each dimension, and the encoding algorithm to be used, such as hash mapping or numerical normalization. Finally, the central control node broadcasts the generated adaptive encoding strategy instruction to all registered edge acquisition probes through a dedicated configuration distribution message queue, and simultaneously pushes the version number of the instruction to the decoding configuration module on the central server for caching.

[0030] For example, during a promotional campaign, the monitoring module of the central control node acquires a set of global operational status parameters at time T. First, the system detects that the average CPU utilization of the server cluster has reached 92%, i.e., the system load rate. Secondly, the system query revealed in the database that the core product SKU "Phone-X123" was currently in a flash sale, i.e., a promotional event. Finally, the system parses the API call headers from the third-party platform and finds that only 500 of the 10,000-call limit remain within the current minute period. The system then processes these parameters to calculate a comprehensive system status score. Normalized system load Calculated as Interface flow control proximity Calculated as Then, the preset weights are substituted. Calculations were performed to obtain The calculated result of 0.95 is greater than the preset scene switching threshold of 0.85, so the strategy generator is triggered. It matches the high-concurrency flash sale scene rules and generates a new adaptive encoding strategy instruction. In the JSON content of this instruction, the field extraction rules are simplified to ["$.userId","$.items[0].sku","$.timestamp"], where $ represents the root directory of the JSON data to capture the core fields required for anti-fraud; the vector space definition specifies a three-dimensional vector structure, with the first dimension performing 1024-bucket hash encoding on the user ID, the second dimension performing table lookup encoding on the product SKU, and the third dimension normalizing the millisecond-level timestamp. Finally, this adaptive encoding strategy instruction containing the new rules is sent to all edge acquisition probes through the Kafka message bus, and its central decoding module also loads the corresponding decoding logic synchronously, thus completing a dynamic adjustment of the strategy.

[0031] In one embodiment of the present invention, the edge acquisition module is configured to perform the following steps: By using edge acquisition probes to monitor order generation events on e-commerce platforms in real time, the original order messages containing full metadata are captured; the field extraction rules in the adaptive encoding strategy instructions are parsed, and the path expressions are applied to the original order messages; unnecessary information is filtered out and numerical values ​​or strings that match the current business scenario are selected to form a target key field set whose structure changes with the strategy.

[0032] Specifically, an edge acquisition probe, acting as a resident lightweight service process, captures newly generated raw order messages containing full metadata in real time by subscribing to real-time messaging services provided by the e-commerce platform or listening to webhook events. It's important to note that an edge acquisition probe is a software agent deployed on a server or cloud function close to the data source. Its role is to preprocess the data before it leaves the original platform. The raw order message, on the other hand, is a complete record of order data generated by the e-commerce platform system without any modification. It is typically in JSON or XML format and contains dozens or even hundreds of fields, such as user information, product details, and price details.

[0033] Upon receiving a new raw order message, the edge acquisition probe immediately reads the adaptive encoding strategy instruction from its local memory. Subsequently, the edge acquisition probe invokes its built-in JSON parsing module to parse and filter the raw order message according to the field extraction rules defined in the adaptive encoding strategy instruction. Each JSONPath expression in the field extraction rules is sequentially applied to the tree-like data structure of the raw order message, precisely locating and extracting numerical values ​​or strings directly related to the current business scenario. All successfully extracted data fields are aggregated to form a temporary, structured target key field set. This target key field set is a streamlined dataset containing only core decision-making information, obtained after filtering by the field extraction rules. Its data structure can be a set of key-value pairs or an ordered list, directly serving as input for the next encoding process. The specific composition of this target key field set, including the number and types of fields it contains, is entirely determined by the currently effective adaptive encoding strategy instruction, thereby achieving dynamic adaptation to decision-making needs under different operational states. This process completes preliminary data cleaning and dimensionality reduction at the source of data generation, filtering out a large amount of unnecessary redundant information.

[0034] For example, the edge acquisition probe deployed on the third-party platform has loaded the adaptive encoding strategy instruction generated in the aforementioned example into memory. The field extraction rule contained in this instruction is a set of JSONPath expressions, specifically ["$.userId","$.items[0].sku","$.timestamp"]. At time T+1ms, the probe receives a raw order message for a new order through the message queue. Its content is a JSON object: {"orderId":"JD-20230315-001","userId":"user-abc-987","timestamp":1678886400123,"items":[{"sku":"Phone-X123","price":5999,"quantity":1},{"sku":"Case-001","price":99,"quantity":1}],"shippingAddress":"...","paymentMethod":"..."}. The edge acquisition probe immediately starts the processing flow. First, it applies the first rule "$.userId" to successfully extract the string value "user-abc-987" from the original order message. Next, it applies the second rule "$.items[0].sku", which points to the sku field of the first element of the items array, thus extracting the string value "Phone-X123", while ignoring the information for the second item in the array, "Case-001". Finally, it applies the third rule "$.timestamp" to extract the long integer value 1678886400123. Ultimately, these three extracted values ​​are combined into a target key field set, which is represented in memory as an ordered set: ["user-abc-987","Phone-X123",1678886400123]. This target key field set is then passed to the next step for further processing.

[0035] In one embodiment of the present invention, the edge encoding module is configured to perform the following steps: Traverse the target key field set, perform hash mapping processing on discrete data, and perform normalization processing on continuous numerical data; assemble the processed data into a one-dimensional array according to the dimensional arrangement order defined by the adaptive encoding strategy instruction; insert the version identifier bit corresponding to the adaptive encoding strategy instruction into the head of the one-dimensional array, and encapsulate it to generate a variant order feature vector.

[0036] Specifically, the edge acquisition probe first parses the loaded adaptive encoding strategy instructions again to obtain the vector construction logic defined therein. This logic includes the specific encoding algorithm to be used for each key field and their dimensional order in the final vector. Next, the probe's execution thread iterates through each data item in the target key field set and generates a modified order feature vector according to the following logic. : For data items determined to be discrete data, such as user IDs or product SKU strings, the probe calls a pre-defined hash mapping function library, such as the MurmurHash3 algorithm, to convert them into integer indices within a fixed range. The hash mapping calculation formula is as follows:

[0037] in, The first feature vector representing the modified order feature vector Values ​​of each dimension Represents the first in the target key field set Each field value. This is the preset number of hash buckets for discrete data, typically 1024. Choosing 1024 as the number of buckets here is a balance between feature collision rate and vector sparsity: for millions of user IDs, although 1024 buckets will cause hash collisions, this level of information loss is acceptable in high-load initial screening scenarios, and can significantly reduce transmission and computing overhead.

[0038] For data items determined to be continuous numerical data, such as order timestamps or payment amounts, the probe performs normalization, linearly mapping them to a floating-point range of 0 to 1. The normalization calculation formula is as follows:

[0039] in, and These are the minimum and maximum values ​​of the normalization interval set for continuous data. They are typically determined based on large-scale statistical analysis of historical data or boundary conditions of the business scenario, such as the start and end times of a specific promotional activity window. To prevent extremely small smoothing coefficients with a denominator of zero, such as 10 -6 In particular, to ensure the robustness of the system, when the actual collected values... Exceeding the preset range In this embodiment, a truncation strategy is adopted: if ,but ;like ,but This is to prevent numerical overflow from causing abnormal model calculations.

[0040] After all fields have been encoded and converted, the probe organizes the generated values ​​into a one-dimensional array according to the dimensional order specified in the adaptive encoding strategy instruction, thus forming a dynamically dimensional modified order feature vector. It should be noted that the modified order feature vector is a numerical vector; its number of dimensions, the value range of each dimension, and its business meaning are dynamically defined by the currently effective adaptive encoding strategy instruction, rather than being fixed. To ensure that the central processing unit can decode correctly, the probe extracts its unique version identifier from the adaptive encoding strategy instruction. This identifier is used to associate the vector with a unique string or integer generated by the strategy, and is used as header information, encapsulated together with the generated modified order feature vector into a data packet. Finally, the edge acquisition probe publishes this data packet as the message body to a preset message topic via a lightweight message queue protocol, such as MQTT or AMQP, to achieve low-latency delivery to the central processing unit.

[0041] For example, the edge detection probe continues to process the sample target key field set, namely ["user-abc-987","Phone-X123",1678886400123]. The probe simultaneously reads the adaptive encoding policy instruction, whose version identifier is "Policy-FS-v2," and specifies the construction logic for the three-dimensional vector. First, it processes the first field, "user-abc-987," which requires a 1024-bucket hash mapping. The probe calls the MurmurHash3 function to calculate the hash value of this string, assuming the result is 3141592653, and then performs a modulo operation to obtain... This value becomes the first dimension of the vector. Next, the second field, "Phone-X123," is processed, and the instruction requires table lookup encoding. The probe accesses the pre-loaded SKU-index mapping dictionary, finding that the index corresponding to "Phone-X123" is 58, which becomes the second dimension of the vector. Then, the third field, 1678886400123, is processed, and the instruction requires normalization within the flash sale's time window. Based on the system configuration, the start time of this flash sale... The end time is 1678886400000ms. The time was 1678886460000ms. The probe underwent normalization calculations. This value becomes the third dimension of the vector. The probe then assembles these three values ​​sequentially to generate the core part of the modified order feature vector: [717,58,0.00205]. Finally, the probe encapsulates the version identifier "Policy-FS-v2" along with this vector into a data object, for example, {"version":"Policy-FS-v2","vector":[717,58,0.00205]}, and publishes it to a topic named OrderVectors via the MQTT protocol for the central processing unit to receive.

[0042] In one embodiment of the present invention, the receiving and loading module is configured to perform the following steps: The central processing unit parses the version identifier bit in the header of the modified order feature vector; indexes the model file path corresponding to the version identifier bit in the pre-set model configuration registry; loads and instantiates the target decoding head model from the pre-trained multimodal decoder library, the target decoding head model being configured with a feature mapping matrix adapted to the current vector structure.

[0043] Specifically, the listening module of the central processing unit acts as a consumer of the message queue, continuously pulling data packets sent by the edge acquisition probes from the lightweight message queue. In this embodiment, the central processing unit is a distributed computing system deployed in the cloud or data center, responsible for receiving and processing all data from the edge. When a new data packet arrives, the central processing unit first deserializes it, extracting the variant order feature vector and its version identifier bit in the header. Subsequently, the engine uses the version identifier bit as a lookup key to search in its internally maintained model configuration registry. This registry stores the mapping relationship between all defined adaptive encoding strategy instruction versions and the corresponding target decoding header model file paths.

[0044] Once a match is successful, the engine's resource scheduler indexes and loads the corresponding target decoder head model file from a centralized, pre-trained multimodal decoder library based on the found file path. The pre-trained multimodal decoder library is a model repository, such as a file system directory, cloud storage bucket, or model management platform, storing multiple pre-trained models for different business scenarios and encoding strategies. The target decoder head model is a pre-trained machine learning model, such as a feedforward neural network containing fully connected layers. Its core function is to inversely map the modified order feature vectors represented in low-dimensional space back to a high-dimensional feature space with clear business meaning. This loading process involves reading the model's binary file, such as an ONNX format file, into memory and instantiating it as an executable inference object. This instantiated target decoder head model internally contains a feature mapping matrix that is the inverse of its corresponding encoding strategy. The feature mapping matrix is ​​a set of weight parameters obtained through offline training within the model, and its values ​​determine the transformation logic from compressed vectors to intermediate features. Finally, the central processing unit passes the memory address of the received modified order feature vector to the input interface of the newly activated target decoding head model, thereby establishing a data channel between the two and preparing for the next step of semantic restoration processing.

[0045] For example, the central processing unit receives a message from an MQTT topic named OrderVectors, containing a sample JSON packet: {"version":"Policy-FS-v2","vector":[717,58,0.00205]}. First, the engine's parsing module extracts the version identifier "Policy-FS-v2" and the variant order feature vector as [717,58,0.00205]. Next, the engine queries its internal model configuration registry using "Policy-FS-v2" as the key. This registry is a hash table structure, with one record being "Policy-FS-v2":" / models / v2 / flash_sale_decoder.onnx". The query returns the storage path of the model file, " / models / v2 / flash_sale_decoder.onnx". Subsequently, the engine's resource scheduler instructs the model loading service to load the file corresponding to this path from the pre-trained multimodal decoder library. The model loading service reads the ONNX file from the local file system, deserializes it, and constructs a target decoding head model instance for high-concurrency flash sale scenarios in memory. This model instance contains a pre-trained feature mapping matrix for decoding the 3D vector structure. Finally, the central processing unit uses the modified order feature vector [717,58,0.00205] as a parameter to call the inference interface of this model instance, thereby establishing a data flow channel and completing this step.

[0046] In one embodiment of the present invention, the decoding and restoration module is configured to perform the following steps: The modified order feature vector is input into the target decoding head model. The hybrid feature mapping layer is used to embed and look up the discrete dimensions and perform linear projection on the continuous dimensions. The projection results are concatenated and restored to intermediate features with business semantics. The intermediate features are projected onto a unified high-dimensional feature space and the dimensions not collected by the current strategy are filled with default values ​​to generate a decision feature tensor.

[0047] Specifically, the inference module of the central processing unit takes the modified order feature vector from the established data channel as input and feeds it into the activated target decoding head model for forward propagation computation. During this process, the target decoding head model employs a hybrid feature mapping mechanism: For discrete index dimensions in the vector, such as the hashed user ID, the model performs a lookup operation through the embedding layer to map the scalar index to a dense real number vector, for example, mapping the index value to a 16-dimensional floating-point vector; For continuous numerical dimensions in the vector, such as normalized time, the model performs weighted projection through linear layers.

[0048] The two features are then concatenated and processed through an activation function to restore the intermediate features with preliminary business semantics. The mathematical expression of this restoration process is shown in the following formula:

[0049] in, Represents intermediate characteristics; and These represent the discrete and continuous parts of the input vector, respectively. This represents the lookup function for the pre-trained embedding matrix; and These are the weight matrix and bias vector of the linear layer, respectively; This represents a non-linear activation function, such as ReLU.

[0050] Subsequently, the intermediate features are fed into the model's feature alignment module. This module, based on the target decoder model's metadata configuration (Manifest file), which defines the index mapping between each dimension of the intermediate features and each dimension of the standard decision space, identifies the specific business meaning of each dimension in the intermediate features and fills these valid feature values ​​into the corresponding positions of the predefined full-dimensional standard vector template. For dimensions in this template that are not filled due to the current encoding strategy, the module uses preset default values ​​for completion. After this alignment and completion process, a decision feature tensor with fixed dimensions and a uniform format is generated. The decision feature tensor is a high-dimensional numerical vector with fixed dimensions. Its number of dimensions is derived from business requirement analysis and is a superset that can cover all key features in all business scenarios. Its function is to shield the vector structure differences caused by different adaptive encoding strategy instructions at the input end, providing a unified input interface for all subsequent classification models.

[0051] For example, the central processing unit continues to process the variant order feature vector of the example. This data is then input into the loaded "flash_sale_decoder.onnx" target decoding head model. First, the semantic reconstruction process is performed.

[0052] The model configuration is defined as follows: the first dimension -717 and the second dimension -58 are discrete indices, and the third dimension -0.00205 is a continuous value.

[0053] Embedded lookup process: Using the built-in embedding matrix, index 717 is mapped to a vector. Index 58 is mapped to a vector ; Linear projection: The continuous value 0.00205 is multiplied by the weight 10 and the bias 0.1, resulting in 0.1205; Concatenation and activation: Concatenate the above results to obtain After processing with the ReLU activation function, assuming all values ​​are positive, the intermediate features are obtained. .

[0054] Next, spatial projection and completion are performed. For example, from the model's metadata, we know that these 5 intermediate features correspond to the 5th dimension—user hash feature, the 22nd dimension—product index feature, the 3rd dimension—time sensitivity feature, the 6th dimension—user-product cross feature, and the 4th dimension—timestamp correction feature—in the standard 128-dimensional decision feature tensor. The engine then creates a vector of length 128, with all elements initialized to 0. Subsequently, it projects the intermediate features... The values ​​are filled into the specified positions. Finally, this 128-dimensional vector containing 5 non-zero values ​​and 123 zero values ​​is the generated normalized decision feature tensor, which is passed to the next step for processing.

[0055] In one embodiment of the present invention, the classification calculation module is used to perform the following steps: The decision feature tensor is input into the primary fast response classifier to calculate the basic priority probability of the order; it is then determined whether the basic priority probability falls within the preset fuzzy interval; if it falls within the fuzzy interval, the deep neural network model is triggered to perform correlation analysis on the decision feature tensor, and the output values ​​of each level of the model are combined to generate intelligent classification result labels.

[0056] Specifically, the system first receives a standardized decision feature tensor and feeds it into a preloaded primary fast-response classifier. In this embodiment, the multi-level classification network is a decision system composed of at least two classification models with different computational complexities connected in series or parallel. The primary fast-response classifier, as the first level of the network, such as a LightGBM-based model, is designed with speed as its priority, used to handle the vast majority of regular orders. It performs high-speed inference on the input decision feature tensor and outputs a probability representing the basic priority of the order. The scalar value is between 0 and 1. A higher value indicates that the order needs to be prioritized or that there is a greater possibility of risk.

[0057] Then, this base priority probability is compared with a preset fuzzy interval. The preset fuzzy interval is determined by a lower threshold. and upper limit threshold The defined closed interval, typically set as [0.2, 0.8], is based on the following logic: orders with a probability below 0.2 are considered highly likely to be safe or low-priority, while orders above 0.8 are considered highly likely to be risky or high-priority. These two types of orders have sufficiently high confidence levels, requiring no additional computational effort. Orders between 0.2 and 0.8 fall into an "ambiguous" gray area with a higher risk of misjudgment, necessitating a more complex model for judgment. If the probability value falls outside the fuzzy interval (extremely high or low), a preliminary classification conclusion is generated directly based on that probability. Conversely, if the basic priority probability value falls inside the fuzzy interval, the order is considered a complex case that is difficult to judge quickly, triggering a secondary refined analysis by a deep neural network model. In this case, the decision feature tensor is further passed to a deep neural network model, such as a multilayer perceptron, as the second level of the network. This deep model performs deeper correlation analysis and feature cross-calculation on the decision feature tensor, outputting a vector containing multiple dimensions, each corresponding to a different risk type probability.

[0058] Finally, the system's decision fusion module integrates the output of the primary classifier or the corrected output of the deep model to generate structured intelligent classification result labels. These labels are standardized data structures, such as JSON, whose internal fields explicitly include the processing priority assigned to the order, the determined risk type, and the corresponding handling suggestions automatically generated by the system. The decision logic and formula for label generation are expressed as follows:

[0059] in, The labels represent the final generated intelligent classification results; This represents the base priority probability output by the primary classifier; Represents the preset fuzzy judgment range; This represents a simple set of threshold mapping rules for the initial classification results, for example: if Then it is directly marked as risky or high priority; This represents a complex set of multidimensional decision rules for the output vectors of deep networks, such as logic that requires a comprehensive judgment combining both the fraud probability dimension and the inventory tightness dimension. The decision feature tensor representing the input; The inference function representing a deep neural network model; Right now It is a tag generation function that constructs the final tag content based on the input probability information and the corresponding rule set.

[0060] For example, the classification computation module of the central processing unit receives a 128-dimensional decision feature tensor generated in the example. First, this tensor is input into a loaded LightGBM primary fast-response classifier. After model inference, the output is the base priority probability. The probability is 0.65. Next, this probability value is compared with a preset fuzzy interval [0.2, 0.8]. Since 0.65 falls within this interval, it is determined that this order requires in-depth analysis, triggering the second-level classification process. Subsequently, this 128-dimensional decision feature tensor is again used as input and passed to a pre-loaded multilayer perceptron deep neural network model. After performing inference calculations, the deep model outputs a three-dimensional risk probability vector: [0.75, 0.1, 0.05], where the three dimensions represent the probabilities of fraud risk, inventory risk, and logistics risk, respectively. Finally, the decision fusion module parses this risk probability vector. Since the probability of fraud risk (0.75) in the first dimension is much higher than other risks and exceeds the preset fraud judgment threshold of 0.7, this module generates an intelligent classification result label accordingly. The final generated JSON format label is: {"priority":"Highest","risk_type":"SuspectedFraudulentOrder","suggestion":"Suspend orderand flag for manual review"}. This label is then pushed to downstream business systems.

[0061] In one embodiment of the present invention, the feedback and update module is used to perform the following steps: The system analyzes the labels of the intelligent classification results to identify risk types and processing priorities; it performs inventory pre-locking operations on orders marked as first priority; it performs interception and suspension operations on orders marked as having specific risk types; and for orders marked as requiring in-depth verification, it initiates a full data asynchronous pull-back thread to obtain complete order details from the source.

[0062] Specifically, the business execution engine in the central processing unit first subscribes to and receives the intelligent classification result tags. The business execution engine is the software module responsible for translating classification decisions into specific business system actions. Upon receiving the tags, the engine parses the internal fields and routes them to different execution logics according to preset business rules. For example, if the risk type in the intelligent classification result tag is marked as abnormal or fraudulent, the engine will invoke an interception and suspension operation, that is, by sending a request to the API interface of the Order Management System (OMS), suspending the order process and placing it in a pending queue. If the tag indicates first priority (high priority), an inventory pre-locking operation may be performed, that is, reserving product inventory for a specific order in the inventory management system but not actually shipping the goods. If the tag indicates that deep verification is required, the business execution engine will instantiate and start a full data asynchronous pull-back thread. This thread is a non-blocking data acquisition process running in the background. Based on the order ID and source platform information, it asynchronously requests the complete original order message from the corresponding e-commerce platform through a standard API interface and persists the acquired data to the backend database for manual review.

[0063] While executing business control logic, the engine's monitoring plugin records the end-to-end latency of this order processing as processing timeliness, and the CPU time consumed by the classification model inference as system resource consumption data. These statistics are sent to the time-series database in real time. An independent statistical analysis service periodically aggregates and calculates the data in the time-series database, updating indicators such as system load rate and risk order ratio in the global operational status parameters. To avoid system oscillations, the parameter updates employ a weighted smoothing mechanism, the formula of which is as follows:

[0064] in, This represents the system load rate after the update at time point t; It is the instantaneous CPU load value collected by the monitoring module at the current moment; It is the system load rate calculated in the previous update cycle; This is the exponential smoothing coefficient, typically 0.2, used to control the weighting of historical data on the assessment of the current state. By setting a smaller value... This value can filter out instantaneous spikes and fluctuations, preventing the policy generator from frequently switching policies due to occasional network jitter, thus ensuring system stability.

[0065] These updated global operational status parameters are fed back to the central control node of the monitoring and strategy generation module. When the central control node detects that any updated parameter, such as the system's average processing time or system load rate, exceeds the preset performance threshold for a continuous period of time, such as the proportion of orders triggering the deep model exceeding 20% ​​in the past 5 minutes, it will trigger the regeneration and issuance of a new adaptive coding strategy instruction, thereby forming a complete closed-loop adaptive control system.

[0066] For example, the business execution engine receives the example intelligent classification result label {"priority":"Highest","risk_type":"SuspectedFraudulentOrder","suggestion":"Suspend orderand flag for manual review"}, associated with the order ID "JD-20230315-001". First, the engine parses the risk_type as "SuspectedFraudulentOrder" and triggers an interception and suspension operation. It sends an HTTP PUT request to the order management system's API endpoint, setting the status to SUSPENDED. Simultaneously, based on the content of the suggestion field, the engine starts a full data asynchronous pull-back thread. Next, the engine's monitoring plugin records the processing time as 150ms and system resource consumption as 0.05 CPU-seconds. This data is sent to the statistical analysis service. This service reads the current instantaneous CPU load. The system load rate is 92%, and the system load rate for the previous cycle is obtained from the internal state. The smoothing coefficient is 88%. The service calculates the updated system load rate: The updated 88.8% was written back to the global operational status parameters. Finally, the central control node read this updated load value and compared it with the preset policy adjustment trigger threshold of 90%. Since 88.8% is less than 90%, no new adaptive coding policy instruction was generated in this update, and the system continued to operate under the current policy, completing the closed-loop feedback.

[0067] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A management system that supports real-time aggregation and intelligent classification of orders from multiple stores, characterized in that, include: The monitoring and strategy generation module is used to obtain global operational status parameters of the multi-channel e-commerce system and generate adaptive coding strategy instructions based on the preset business scenario model. The edge acquisition module uses edge acquisition probes to capture the original order message and extracts the target key field set according to the adaptive encoding strategy instructions; The edge encoding module, based on the vector construction logic defined by the adaptive encoding strategy instruction, performs structured encoding on the target key field set to generate a variant order feature vector; The receiving and loading module uses the central processing unit to receive the variant order feature vector and activates the target decoding head model that matches the current strategy; The decoding and restoration module uses the target decoding head model to perform semantic restoration and spatial projection on the variant order feature vector, generating a standardized decision feature tensor. The classification calculation module inputs the decision feature tensor into a multi-level classification network, calculates the business attributes and risk level of the order, and outputs intelligent classification result labels. The feedback and update module responds to the intelligent classification result labels, executes business control logic, and updates global operational status parameters to form a closed-loop control.

2. The management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The monitoring and policy generation module is used to perform the following steps: By monitoring the system load rate, the promotional activity status of specific products, and the flow control threshold of the platform interface through the central control node, global operational status parameters are generated. Input the global operational status parameters into the preset strategy generator to calculate the overall system status score; When the overall system status score exceeds the preset threshold, an adaptive encoding strategy instruction containing field extraction rules and vector space definition is generated and sent to the edge acquisition probe.

3. The management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The edge acquisition module is used to perform the following operations: Use edge acquisition probes to monitor order generation events on e-commerce platforms in real time and capture raw order messages containing full metadata; Parse the field extraction rules in the adaptive encoding strategy instruction and apply the path expression to the original order message; Filter out unnecessary information and select numerical values ​​or strings that match the current business scenario to form a target key field set with a variable length as the strategy changes.

4. The management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The edge coding module is used to perform the following steps: Iterate through the target key field set, perform hash mapping processing on discrete data, and perform normalization processing on continuous numerical data; The processed data is assembled into a one-dimensional array according to the dimensional arrangement order defined by the adaptive encoding strategy instructions; A version identifier corresponding to the adaptive encoding strategy instruction is inserted into the header of a one-dimensional array, and the modified order feature vector is generated by encapsulation.

5. A management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 4, characterized in that, The receive and load module is used to perform the following steps: The central processing unit parses the version identifier bit in the header of the modified order feature vector; The model file path corresponding to the version identifier is indexed in the preset model configuration registry; Load and instantiate the target decoder head model from the pre-trained multimodal decoder library. The target decoder head model is configured with a feature mapping matrix that is adapted to the current vector structure.

6. A management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The decoding and restoration module is used to perform the following steps: The modified order feature vector is input into the target decoding head model, and the hybrid feature mapping layer is used to perform table embedding for discrete dimensions and linear projection for continuous dimensions. The projection results are stitched together and restored to intermediate features with business semantics; The intermediate features are projected onto a unified high-dimensional feature space, and the dimensions not collected by the current strategy are filled with default values ​​to generate a decision feature tensor.

7. A management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The classification calculation module is used to perform the following steps: Input the decision feature tensor into the primary fast response classifier to calculate the basic priority probability of the order; Determine whether the basic priority probability falls within a preset fuzzy range; If the value falls within the fuzzy range, the deep neural network model is triggered to perform correlation analysis on the decision feature tensor, and the output values ​​of each level of the model are combined to generate intelligent classification result labels.

8. A management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The feedback and update module is used to perform the following steps: Analyze the labels of intelligent classification results to identify risk types and processing priorities; Perform inventory pre-locking on orders marked as first priority; For orders marked with a specific risk type, execute an interception and suspension operation; For orders marked as requiring in-depth verification, initiate an asynchronous full data pull-back thread to obtain complete order details from the source.

9. A management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 1, characterized in that, The feedback and update module is also used to perform the following sub-steps to update global operational status parameters: Statistical analysis of order processing time and system resource consumption data; The current statistical data and historical data are weighted and merged using a smoothing coefficient to update the global operational status parameters; The updated global operational status parameters are compared with the preset scene switching threshold. If the threshold is exceeded, a new adaptive coding strategy instruction is generated.

10. A management system supporting real-time aggregation and intelligent classification of orders from multiple stores according to claim 2, characterized in that, The specific methods for monitoring each parameter include: The utilization rate of the central processing unit is obtained by reading the server performance counters as the system load rate; The current enumeration value of the product SKU is obtained by performing a database query operation as the status of the promotional activity; The remaining call count in the response header returned by the e-commerce platform API is used as the flow control threshold for the platform interface.

Images