Dial testing method, device and equipment for distribution channel compliance and storage medium

Abstract

The invention provides a distribution channel compliance dial testing method, device and equipment and a storage medium, and relates to the technical field of dial testing, and the method comprises the steps: carrying out the analysis of order snapshot data in a current time window of a distribution channel system, and obtaining a multi-modal feature vector; detecting whether the distribution channel system is illegal or not according to the multi-modal feature vector; when it is detected that the distribution channel system is illegal, generating a hash value according to the order snapshot data and the illegal information of the distribution channel system, and uploading the hash value to a block chain node for evidence storage to form real-time evidence storage information; and obtaining an electronic signature and a timestamp returned by the block chain node for the real-time evidence storage information. According to the scheme of the invention, compliance detection can be performed in real time in a distribution channel acceptance process, so that illegal behaviors can be prevented in time.

Description

Technical Field

[0001] This invention relates to the field of dialing testing technology, and specifically to a dialing testing method, apparatus, equipment, and storage medium for testing the compliance of distribution channels. Background Technology

[0002] The testing methods for distribution channel compliance are used to automate compliance checks on elements of distribution channel pages, such as product information, compliance logos, interactive components, and processing flows, such as qualification review, order processing, and permission verification. Existing testing methods for distribution channel compliance primarily rely on pre-set test scripts and templates, making it difficult to cope with dynamic changes to distribution channel pages. Furthermore, they focus on post-event monitoring and cannot perform real-time compliance checks during the distribution channel processing flow, thus failing to promptly prevent violations. Summary of the Invention

[0003] At least one embodiment of the present invention provides a method, apparatus, device and storage medium for testing the compliance of distribution channels, which solves the problem in the prior art that compliance testing cannot be performed in real time during the acceptance process of distribution channels, making it difficult to prevent violations in a timely manner.

[0004] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0005] In a first aspect, embodiments of the present invention provide a method for testing the compliance of distribution channels, including:

[0006] The order snapshot data within the current time window of the distribution channel system is parsed to obtain multimodal feature vectors;

[0007] Based on the multimodal feature vector, detect whether there are any violations in the distribution channel system;

[0008] If a violation is detected in the distribution channel system, a hash value is generated based on the order snapshot data and the violation information of the distribution channel system, and the hash value is uploaded to the blockchain node for evidence storage, forming real-time evidence storage information;

[0009] Obtain the electronic signature and timestamp returned by the blockchain node for the real-time evidence storage information.

[0010] Optionally, the method for testing the compliance of distribution channels further includes:

[0011] Retrieve order lists from the customer relationship management system at a preset frequency;

[0012] Based on the order list, obtain multiple initial order snapshot data from the distribution channel system;

[0013] Within the current time window, updated and / or changed order snapshot data are filtered from multiple initial order snapshot data.

[0014] Optionally, the method for testing the compliance of distribution channels includes parsing order snapshot data within the current time window of the distribution channel system to obtain a multimodal feature vector, including:

[0015] The text feature vector is obtained by parsing the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system.

[0016] Image feature vectors are obtained by parsing the compliance identifiers in the order snapshot data;

[0017] Semantic parsing is performed on the contract terms in the order snapshot data to obtain semantic feature vectors;

[0018] The text feature vector, the image feature vector, and the semantic feature vector are concatenated to obtain a multimodal feature vector.

[0019] Optionally, the method for testing the compliance of distribution channels includes parsing the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system to obtain text feature vectors, including:

[0020] By monitoring the Document Object Model (DOM) node tree, it can be determined whether the qualification documents and page text in the order snapshot data of the distribution channel system have changed;

[0021] If the qualification document and the page text are determined to have changed, the qualification document and the page text are parsed to obtain text feature vectors.

[0022] Optionally, the method for detecting compliance of the distribution channel includes, based on the multimodal feature vector, detecting whether the distribution channel system violates regulations, including:

[0023] Based on the multimodal feature vectors and anomaly detection model, detect whether there are any violations in the distribution channel system;

[0024] The anomaly detection model includes a supervised learning module, a semi-supervised learning module, and a reinforcement learning module. The supervised learning module pre-builds a compliance judgment benchmark and uses it to determine whether the multimodal feature vector is in violation based on the compliance judgment benchmark, thereby obtaining a judgment result. The semi-supervised learning module is used to identify the suspected violation type corresponding to the multimodal feature vector, thereby obtaining an identification result. The reinforcement learning module is used to obtain a compliance score for the distribution channel system based on the judgment result and the identification result, and to detect whether the distribution channel system is in violation based on the compliance score.

[0025] Optionally, the method for testing the compliance of distribution channels may further include one of the following:

[0026] The risk level of the distribution channel system is obtained based on the compliance score. Based on the risk level and the risk diffusion trend prediction model, the risk diffusion trend information of the distribution channel system is obtained. Corresponding early warning processing is then performed based on the risk diffusion trend information.

[0027] The distribution channel system is mapped to a color in the risk warning heatmap based on the compliance score, and the distribution channel system is displayed in the risk warning heatmap based on the mapped color.

[0028] Optionally, the method for testing the compliance of distribution channels further includes:

[0029] Based on a pre-constructed violation transmission chain, historical violation evidence related to the distribution channel system is obtained; the violation transmission chain is constructed based on distributor information, qualification documents, and violation types extracted from historical order snapshot data.

[0030] Based on the historical evidence of violations and the decision data determined by the anomaly detection model that the multimodal feature vector contains violations, violation information of the distribution channel system is obtained.

[0031] Optionally, the method for testing the compliance of the distribution channel includes uploading the hash value to a blockchain node for notarization, forming real-time notarization information, including:

[0032] Obtain historical evidence information associated with the aforementioned violation information;

[0033] Triggering a notarization transaction according to preset business rules, the hash value and the historical notarization information are uploaded to the blockchain node for notarization, forming real-time notarization information.

[0034] Secondly, embodiments of the present invention also provide a device for testing the compliance of distribution channels, comprising:

[0035] The parsing module is used to parse the order snapshot data within the current time window of the distribution channel system to obtain multimodal feature vectors;

[0036] The detection module is used to detect whether there are any violations in the distribution channel system based on the multimodal feature vector.

[0037] The evidence storage module is used to generate a hash value based on the order snapshot data and the violation information of the distribution channel system when a violation is detected in the distribution channel system, and upload the hash value to the blockchain node for evidence storage to form real-time evidence storage information;

[0038] The acquisition module is used to acquire the electronic signature and timestamp returned by the blockchain node for the real-time evidence storage information.

[0039] Thirdly, embodiments of the present invention also provide a distribution channel compliance testing device, comprising: a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the processor executes the program or instructions to implement the distribution channel compliance testing method as described in the first aspect.

[0040] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the distribution channel compliance testing method as described in the first aspect.

[0041] Fifthly, embodiments of the present invention also provide a computer program product, including computer instructions, which, when executed by a processor, implement the distribution channel compliance testing method as described in the first aspect.

[0042] Compared with existing technologies, this invention provides a method, apparatus, device, and storage medium for testing the compliance of distribution channels. It parses order snapshot data within the current time window of the distribution channel system to obtain a multimodal feature vector. Based on the multimodal feature vector, it detects whether the distribution channel system has any violations. If a violation is detected, it generates a hash value based on the order snapshot data and the violation information of the distribution channel system, and uploads the hash value to a blockchain node for notarization, forming real-time notarized information. It obtains the electronic signature and timestamp returned by the blockchain node for the real-time notarized information. Thus, time windows can be divided according to needs, and distribution channel compliance testing can be performed within the current time window to achieve real-time and dynamic compliance testing in the distribution acceptance process, thereby promptly preventing violations. Moreover, in the case of a detected violation in the distribution channel system, the violation information is notarized through blockchain, ensuring the legal validity of the test results. Attached Figure Description

[0043] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0044] Figure 1 This is a flowchart illustrating the distribution channel compliance testing method described in an embodiment of the present invention.

[0045] Figure 2 This is a timing diagram of the order snapshot data collection according to an embodiment of the present invention;

[0046] Figure 3 This is a diagram illustrating the hierarchical storage architecture of order snapshot data as described in an embodiment of the present invention.

[0047] Figure 4 This is a pipeline diagram of the multimodal feature vector extraction described in an embodiment of the present invention;

[0048] Figure 5 This is a structural diagram of the cross-modal attention mechanism provided in an embodiment of the present invention;

[0049] Figure 6 The training loss curve of the joint model provided in the embodiments of the present invention;

[0050] Figure 7 A schematic diagram of the dynamic DOM parsing algorithm provided in an embodiment of the present invention;

[0051] Figure 8 This is a comparison chart of lightweight transmission compression of order snapshot data according to an embodiment of the present invention;

[0052] Figure 9 This is a flowchart illustrating the training process of the anomaly detection model described in an embodiment of the present invention.

[0053] Figure 10 The real-time risk warning heat map provided in the embodiments of the present invention;

[0054] Figure 11 This is a schematic diagram illustrating the construction of a knowledge graph in the distribution field provided in an embodiment of the present invention;

[0055] Figure 12 A blockchain-based evidence storage technology roadmap provided for embodiments of the present invention;

[0056] Figure 13 This is a schematic diagram illustrating the binding of electronic signatures and timestamps according to an embodiment of the present invention;

[0057] Figure 14This is an overall flowchart of the distribution channel compliance testing method described in this embodiment of the invention;

[0058] Figure 15 This is a flowchart illustrating the overall process sequence of the distribution channel compliance testing method described in this embodiment of the invention.

[0059] Figure 16 A partial interpretation heatmap of LIME provided in an embodiment of the present invention;

[0060] Figure 17 This is a global dependency tree analysis diagram of SHAP values ​​provided in an embodiment of the present invention;

[0061] Figure 18 This is a schematic diagram of an edge-cloud task scheduling decision tree provided in an embodiment of the present invention;

[0062] Figure 19 A latency comparison diagram of the edge-cloud collaborative architecture provided in an embodiment of the present invention;

[0063] Figure 20 A flowchart for the compliance testing of package tariffs provided in this embodiment of the invention;

[0064] Figure 21 A schematic diagram illustrating the compliance status of the package ordering process provided in this embodiment of the invention;

[0065] Figure 22 A word cloud diagram for detecting advertising and promotional violation keywords provided in this embodiment of the invention;

[0066] Figure 23 This is a video playback diagram of the value-added service ordering process provided in an embodiment of the present invention;

[0067] Figure 24 This is a timing diagram for abnormal call volume detection provided in an embodiment of the present invention;

[0068] Figure 25 A diagram illustrating the association of fraudulent phone numbers provided in this embodiment of the invention;

[0069] Figure 26 This is a matrix diagram of intelligent work order dispatch rules provided in an embodiment of the present invention;

[0070] Figure 27 This is a histogram comparing work order processing efficiency provided in an embodiment of the present invention.

[0071] Figure 28 This is a diagram illustrating the multi-platform interface adaptation layer architecture provided in this embodiment of the invention.

[0072] Figure 29 This is a browser compatibility test coverage diagram provided in an embodiment of the present invention;

[0073] Figure 30This is a data desensitization technology roadmap provided for embodiments of the present invention;

[0074] Figure 31 This is a diagram of the RBAC (Right-Based Access Control) model for permission management provided in an embodiment of the present invention.

[0075] Figure 32 This is a schematic diagram of a full-link monitoring dashboard provided in an embodiment of the present invention;

[0076] Figure 33 This is a model incremental learning loop diagram provided in an embodiment of the present invention;

[0077] Figure 34 This is a functional matrix diagram of a telecommunications industry customized package provided in an embodiment of the present invention;

[0078] Figure 35 A comparison chart of SaaS subscription and local deployment costs provided for embodiments of the present invention;

[0079] Figure 36 This is a schematic diagram of the module of the distribution channel compliance testing and recommendation device according to an embodiment of the present invention;

[0080] Figure 37 This is a hardware block diagram of the distribution channel compliance testing and recommendation device described in an embodiment of the present invention. Detailed Implementation

[0081] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, the first object can be one or more. Furthermore, the "or" in this invention indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: including A but not B; Scenario 2: including B but not A; Scenario 3: including both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0082] See Figure 1 This invention provides a method for testing the compliance of distribution channels, which can be applied in the telecommunications field.

[0083] Furthermore, the method includes:

[0084] Step 101: Analyze the order snapshot data within the current time window of the distribution channel system to obtain multimodal feature vectors;

[0085] The order snapshot data is a type of multimodal data, and it includes at least screenshots of the order page and order logs. Therefore, parsing the order snapshot data within the current time window of the distribution channel system in this embodiment of the invention is essentially multimodal data parsing to obtain multimodal feature vectors.

[0086] In one embodiment, prior to step 101, the method further includes:

[0087] Retrieve order lists from the customer relationship management system at a preset frequency;

[0088] Based on the order list, obtain multiple initial order snapshot data from the distribution channel system;

[0089] Within the current time window, updated and / or changed order snapshot data are filtered from multiple initial order snapshot data.

[0090] Figure 2 This is a timing diagram illustrating the acquisition of order snapshot data according to an embodiment of the present invention, used to demonstrate the timing logic from the acquisition of the order snapshot data to the generation of the order snapshot data; as follows: Figure 2 As shown, this embodiment of the invention retrieves an order list from the telecommunications customer relationship management (CRM) system at a preset frequency (e.g., 100 times / second). This triggers a snapshot generation engine to capture order page screenshots and order logs from the distribution channel system based on the order list, obtaining multiple initial order snapshot data. Multi-tenant isolation is supported to avoid data confusion. Within the current time window, updated and / or changed order snapshot data are filtered from these initial order snapshot data. Only the changed order snapshot data is transmitted to the distributed file system. The storage format can include HTML, JSON, or screenshots, ensuring data integrity.

[0091] For example, when the promotional information on the page of an e-commerce distribution channel system is updated, this embodiment of the invention can automatically capture the changes and generate new order snapshot data, avoiding repeated collection of static content and reducing traffic consumption.

[0092] Figure 3 This is a diagram illustrating the hierarchical storage architecture of order snapshot data according to an embodiment of the present invention. Figure 3 As shown, the order snapshot data in this embodiment of the invention adopts a data tiered storage strategy. Figure 3 This is used to illustrate the data tiered storage strategy, distinguishing between the storage management of high-frequency hot data and low-frequency cold data.

[0093] exist Figure 3In the storage layer, the hot data layer uses SSDs to store order snapshots from the past 30 days that are accessed frequently, while the cold data layer archives order snapshots from infrequently accessed data using HDDs, with automatic daily backups by the version management module. A compression algorithm component performs Gzip compression (e.g., a compression ratio of 8:1) on the cold data to reduce storage costs.

[0094] For example, historical order snapshot data is automatically compressed when archived, freeing up resources to support the collection of tens of millions of data points per day.

[0095] Therefore, this embodiment of the invention achieves a balance between the real-time performance and storage efficiency of order snapshot data in the distribution channel system through high-frequency collection, incremental filtering, and hierarchical storage. It also supports multi-tenant isolated collection to ensure data security, optimizes storage efficiency through hierarchical storage, and reduces storage costs by using compression technology. Compared with existing manual collection, it improves efficiency, reduces the missed collection rate, and solves the problems of missed collection and lag in existing manual collection. At the same time, it can provide a reliable data source for compliance detection in subsequent steps.

[0096] In one implementation, optionally, step 101 involves parsing the order snapshot data within the current time window of the distribution channel system to obtain a multimodal feature vector, including:

[0097] The text feature vector is obtained by parsing the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system.

[0098] Image feature vectors are obtained by parsing the compliance identifiers in the order snapshot data;

[0099] Semantic parsing is performed on the contract terms in the order snapshot data to obtain semantic feature vectors;

[0100] The text feature vector, the image feature vector, and the semantic feature vector are concatenated to obtain a multimodal feature vector.

[0101] Figure 4 This is a pipeline diagram for extracting multimodal feature vectors according to an embodiment of the present invention, used to illustrate the dynamic processing flow of technology fusion. Figure 4 In this context, modules based on various technologies are deeply integrated with distribution scenarios.

[0102] like Figure 4As shown, this embodiment of the invention uses a text recognition module based on Optical Character Recognition (OCR) technology, referred to as the OCR text recognition module, which not only parses static qualification documents (such as distributor business licenses) in the order snapshot data with high accuracy, but also performs incremental recognition on real-time updated page text (such as promotional activity rules). Combined with DOM change monitoring (the trigger rate can reach 100%), it only processes newly added text fragments (such as the supplementary explanation of "first month half price"), avoiding repeated parsing.

[0103] The image analysis module based on computer vision (CV) technology, referred to as the CV image analysis module, detects compliance marks (such as the telecommunications business operation license icon) in the order snapshot data with high accuracy. At the same time, it identifies hidden violations such as "reuse of expired qualifications" through dynamic comparison technology of image features (such as verification of the validity period watermark of qualification documents with the current time).

[0104] When extracting keywords (such as "prohibited distribution areas" and "commission ratio cap") from the contract terms summarized by the order snapshot data using the semantic understanding module based on Natural Language Processing (NLP) technology, the module can also perform semantic association by combining the distribution industry regulatory library (such as the "Administrative Measures for Telecommunications Business Operation Licenses"). For example, it can automatically associate "vaguely stated service period" with the "suspected false advertising" rule.

[0105] Finally, through the multimodal fusion layer, a cross-modal attention mechanism (e.g., 12-head attention) is used to calculate feature interaction weights. When concatenating the text feature vector, image feature vector, and semantic feature vector into a 1024-dimensional vector, the focus is on strengthening business-related features (e.g., the logical binding of the "limited-time discount" text with the promotional icon and the order submission button), providing contextual feature support for anomaly detection in subsequent steps.

[0106] In this invention, which addresses the challenges of diverse qualification documents (such as PDF contracts and handwritten work orders) and complex business rules in telecommunications distribution scenarios, a cross-modal attention mechanism and joint training model are proposed to achieve deep fusion of text, image, and time-series data, thereby solving the problem of insufficient accuracy in single-modal detection.

[0107] Figure 5 This is a structural diagram of the cross-modal attention mechanism provided in an embodiment of the present invention, used to illustrate the interaction logic between text and image features.

[0108] like Figure 5As shown, for example, the 768-dimensional text feature vector output by BERT and the 2048-dimensional image feature vector output by ResNet50 are respectively input into the cross-modal attention layer, and the interaction weights are calculated through the 12-head attention mechanism to strengthen key features (such as the association between package pricing text and promotional icons).

[0109] For example, when inspecting a telecommunications value-added service page, the model uses an attention mechanism to associate the "first month half price" text with the limited-time icon to accurately identify illegal promotions.

[0110] Figure 6 The training loss curve of the joint model provided in this embodiment of the invention is used to show the improvement of detection accuracy by joint training.

[0111] like Figure 6 As shown, after 50 epochs of model iteration driven by the multi-task loss function, the validation set accuracy increased from 78% to 93%, and the loss value decreased from 2.5 to 0.8, indicating that cross-modal joint modeling can improve detection accuracy. For example, when parsing the real-name authentication page for telecommunications, the joint model simultaneously identifies the clarity of the ID card image and the standardization of text filling, and the accuracy of compliance judgment can be improved by 15% compared with the single modality.

[0112] Therefore, this invention provides a multimodal feature joint modeling method. Through cross-modal attention mechanism and joint training, it realizes collaborative analysis of multiple types of data in telecommunications scenarios, improves detection accuracy, effectively covers multi-source data such as qualification documents, promotional materials, and operation logs, solves the problem of insufficient understanding of complex business rules by existing single modal models, and provides more comprehensive feature support for telecommunications distribution compliance detection.

[0113] In one implementation, optionally, text parsing is performed on the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system to obtain text feature vectors, including:

[0114] By monitoring the Document Object Model (DOM) node tree, it can be determined whether the qualification documents and page text in the order snapshot data of the distribution channel system have changed;

[0115] If the qualification document and the page text are determined to have changed, the qualification document and the page text are parsed to obtain text feature vectors.

[0116] In this embodiment of the invention, a dynamic Document Object Model (DOM) parsing algorithm is proposed to address the frequent dynamic updates of order pages in telecommunications distribution channel systems (e.g., real-time promotions and multi-tenant customized interfaces). This algorithm solves the problem of delayed response to dynamic content in traditional crawling schemes and enables real-time capture of changes to order pages.

[0117] Figure 7 This is a schematic diagram of the dynamic DOM parsing algorithm provided in an embodiment of the present invention, used to illustrate the dynamic DOM change monitoring and incremental rendering logic;

[0118] like Figure 7 As shown, MutationObserver can monitor changes to the DOM node tree in real time, achieving an event capture rate of 100%. It triggers the virtual DOMdiff algorithm to calculate the differences in less than 100ms, and only loads the changed elements into the incremental rendering queue, reducing bandwidth consumption.

[0119] For example, when the "limited-time discount" label on the telecom package page is updated in real time, only the changing DOM nodes are captured to avoid repeatedly loading the entire page and improve data collection efficiency.

[0120] Furthermore, considering the frequent dynamic updates of order pages in telecommunications distribution channel systems (such as real-time promotions and multi-tenant customized interfaces), a lightweight transmission compression technology for the order snapshot data is proposed to solve the problem of high traffic consumption for dynamic content in traditional crawling schemes and achieve efficient transmission.

[0121] Figure 8 This is a comparison chart of lightweight transmission compression of order snapshot data according to an embodiment of the present invention, used to compare the transmission efficiency of the traditional solution and the lightweight solution of the present invention.

[0122] like Figure 8 As shown, traditional solutions require 3 seconds to transmit a complete 5MB page. The lightweight solution in this invention uses WebP image compression and JSON data format to compress page data in order snapshots to 800KB, shortening loading time and significantly improving transmission efficiency. It also supports multi-tenant isolated crawling, ensuring independent collection of page data from different carrier sub-brands and avoiding data confusion.

[0123] Therefore, this invention achieves real-time, low-traffic data collection of dynamic telecommunications pages through dynamic DOM parsing and efficient compression. Compared to traditional solutions, it reduces traffic consumption, improves loading speed, effectively addresses high-frequency page change scenarios in the telecommunications industry, and provides a real-time and complete data foundation for compliance testing in subsequent steps.

[0124] Step 102: Detect whether there are any violations in the distribution channel system based on the multimodal feature vector;

[0125] In one implementation, optionally, step 102, detecting whether there are violations in the distribution channel system based on the multimodal feature vector, includes:

[0126] Based on the multimodal feature vectors and anomaly detection model, detect whether there are any violations in the distribution channel system;

[0127] The anomaly detection model includes a supervised learning module, a semi-supervised learning module, and a reinforcement learning module. The supervised learning module pre-builds a compliance judgment benchmark and uses it to determine whether the multimodal feature vector is in violation based on the compliance judgment benchmark, thereby obtaining a judgment result. The semi-supervised learning module is used to identify the suspected violation type corresponding to the multimodal feature vector, thereby obtaining an identification result. The reinforcement learning module is used to obtain a compliance score for the distribution channel system based on the judgment result and the identification result, and to detect whether the distribution channel system is in violation based on the compliance score.

[0128] In this embodiment of the invention, an anomaly detection model based on artificial intelligence (AI) technology is provided, and therefore this anomaly detection model can also be called an AI anomaly detection model.

[0129] Figure 9 This is a flowchart illustrating the training process of the anomaly detection model according to an embodiment of the present invention, used to demonstrate the closed-loop training logic of the anomaly detection model from data learning to policy optimization.

[0130] like Figure 9 As shown, the anomaly detection model employs multiple learning modes, which complement each other synergistically. The anomaly detection model includes:

[0131] The supervised learning module trains a basic detection model based on millions of labeled violation samples (such as typical cases of "false advertising" and "lack of qualifications") to build an initial compliance judgment benchmark.

[0132] The semi-supervised learning module mines potential risks from massive amounts of unlabeled operational logs (such as user operation trajectories and page interaction records) and identifies suspected violation types (such as abnormal approval paths for a certain type of order) through clustering algorithms, which can improve sample utilization and solve the problem of scarce labeled data.

[0133] The reinforcement learning module introduces a dynamic reward function specific to the distribution scenario to obtain a compliance score. High-risk actions (such as "selling value-added services beyond the scope" and "failing to specify the contract period") are assigned a negative reward of -0.8; while accurate identification of new violations (such as "disguised predation") is assigned a positive reward of +1.2, driving the model to prioritize high-risk scenarios. This dynamic optimization improves the model's convergence accuracy. In the telecommunications package scenario, the accuracy of identifying hidden violations such as "half-price for the first month not prominently displayed" can be increased from 78% to 93%, far exceeding the detection capabilities of existing fixed-rule methods.

[0134] In one embodiment, optionally, the method further includes one of the following:

[0135] The risk level of the distribution channel system is obtained based on the compliance score. Based on the risk level and the risk diffusion trend prediction model, the risk diffusion trend information of the distribution channel system is obtained. Corresponding early warning processing is then performed based on the risk diffusion trend information.

[0136] The distribution channel system is mapped to a color in the risk warning heatmap based on the compliance score, and the distribution channel system is displayed in the risk warning heatmap based on the mapped color.

[0137] Figure 10 The real-time risk warning heat map provided in this embodiment of the invention is used to intuitively present the spatial distribution of risk levels and the real-time warning mechanism, so as to realize the visualization of risks and immediate response.

[0138] like Figure 10 As shown, in this embodiment of the invention, compliance scores (0-100 points) are mapped to a three-color gradient of red (less than 60 points), yellow (60-85 points), and green (greater than 85 points), displaying the distribution of violations in each distribution channel system (such as provincial agents and offline stores) in real time. The red area intuitively marks high-risk points (such as distributors in a certain region frequently using "nationally prohibited" promotional slogans).

[0139] Meanwhile, this invention provides a risk diffusion trend prediction model based on Long Short-Term Memory (LSTM) networks. Combining the risk diffusion trend information output by this model, a second-level warning is triggered for areas where the risk level of the risk diffusion trend information indicates a sudden increase (e.g., a 50% increase in the frequency of violations within 2 hours). The warning is then pushed to the operations team through an interface, reducing the response time from 2 hours for manual inspection to real time, thus achieving "risks appear immediately and warnings are delivered immediately".

[0140] The anomaly detection module achieves a triple breakthrough through the collaborative optimization of multiple learning modes: semi-supervised learning reduces reliance on labeled data, thereby improving sample utilization; reinforcement learning dynamically adapts to the complex rules of distribution scenarios, reducing false alarm rates; and the real-time early warning mechanism compresses risk response time to the minute level, significantly enhancing the ability to identify hidden risks such as "new types of illegal rhetoric" and "cross-channel collaborative violations," providing proactive defense capabilities for distribution compliance detection.

[0141] Therefore, the anomaly detection module in this embodiment of the invention is not a simple superposition of reinforcement learning and supervised learning algorithms, but rather a construction of an intelligent detection system of "dynamic perception - accurate judgment - real-time early warning". Through deep fusion of multiple learning modes, the multimodal feature vectors are used for scenario-based reasoning to generate accurate compliance scores and real-time risk warnings. It is specifically designed to proactively identify suspected violation types (such as hidden promotional language and cross-regional distribution vulnerabilities) that are constantly evolving in the distribution channel system, thus solving the problem of existing static rule engines lagging in responding to emerging risks.

[0142] In one embodiment, optionally, the method further includes:

[0143] Based on a pre-constructed violation transmission chain, historical violation evidence related to the distribution channel system is obtained; the violation transmission chain is constructed based on distributor information, qualification documents, and violation types extracted from historical order snapshot data.

[0144] Based on the historical evidence of violations and the decision data determined by the anomaly detection model that the multimodal feature vector contains violations, violation information of the distribution channel system is obtained.

[0145] In this embodiment of the invention, a knowledge graph in the distribution domain is constructed, which may be referred to as a knowledge graph for short. The knowledge graph includes multiple violation transmission chains, and each violation transmission chain includes at least two transmission nodes.

[0146] Figure 11 This is a schematic diagram of the construction of a knowledge graph in the distribution field provided in an embodiment of the present invention, used to illustrate the association logic between the knowledge graph and business rules.

[0147] Here, combined Figure 11 This illustrates the deep coupling between knowledge graphs and business processes:

[0148] The entity extraction module extracts millions of nodes such as distributor information, qualification documents and violation types from historical order snapshot data, and adds a "dynamic attribute" dimension (such as the validity period of distributor qualifications and the update time of page elements).

[0149] By employing a relationship mining module, basic relationships such as "containment" and "matching" are identified, and violation transmission chains are constructed. Each violation transmission chain includes at least two transmission nodes. For example, a violation transmission chain may include three transmission nodes: "expired qualification", "sales beyond the scope", and "invalid order". The "expired qualification" transmission node points to the "sales beyond the scope" transmission node, and the "sales beyond the scope" transmission node points to the "invalid order" transmission node. Based on this violation transmission chain, the aforementioned knowledge graph can be constructed, thereby improving rule coverage.

[0150] Employing a knowledge reasoning engine, semantic-level compliance verification is achieved through SPARQL queries, with a response time of less than 500ms. For example, when "unqualified distributors selling value-added services" is detected, the system automatically associates the multimodal feature vectors of the corresponding historical order snapshot data (such as screenshots of missing pages of handwritten agreements and page text without distribution scope markings) to form a complete chain of evidence, which can improve efficiency by 5 times compared to existing keyword matching.

[0151] The multimodal parsing engine in this invention achieves contextual understanding of unstructured data through technology fusion and business rule mapping: a 97% text recognition accuracy ensures accurate information extraction, cross-modal association enables semantic compliance verification to cover 95% of industry rules, and the dynamic parsing strategy improves processing efficiency. It provides three-dimensional feature support for anomaly detection, consisting of "element features + business logic + violation propagation", and can accurately identify complex violation scenarios such as "selling after the expiration of qualifications" and "promotional text not matching actual rules".

[0152] Therefore, the multimodal data parsing in this embodiment of the invention is not a simple integration of OCR, CV, and NLP technologies, but rather a deep fusion system of "dynamic perception - cross-modal association - business rule mapping". It achieves full-link parsing from "element recognition" to "business compliance judgment" for unstructured data that changes in real time in the distribution channel system (such as dynamic page elements, heterogeneous qualification documents, and complex contract terms), thus solving the problem of insufficient adaptation of existing single-modal technologies to dynamic scenarios.

[0153] Step 103: If a violation is detected in the distribution channel system, a hash value is generated based on the order snapshot data and the violation information of the distribution channel system, and the hash value is uploaded to the blockchain node for evidence storage to form real-time evidence storage information;

[0154] In one implementation method, optionally, step 103 involves uploading the hash value to a blockchain node for evidence storage, forming real-time evidence storage information, including:

[0155] Obtain historical evidence information associated with the aforementioned violation information;

[0156] Triggering a notarization transaction according to preset business rules, the hash value and the historical notarization information are uploaded to the blockchain node for notarization, forming real-time notarization information.

[0157] In this embodiment of the invention, a blockchain-based evidence storage technology is provided. Figure 12 The blockchain evidence preservation technology roadmap provided for embodiments of the present invention is used to demonstrate the real-time process from evidence generation to on-chain solidification.

[0158] Here, combined Figure 12 This illustrates the deep coupling between the evidence storage process and the entire testing chain:

[0159] A hash generation module is employed. This module does not run independently but is triggered synchronously with the output of the anomaly detection model. When a violation is detected in the distribution channel system (e.g., value-added services do not trigger secondary confirmation), a hash value is immediately generated for the violation evidence, including the order snapshot data and the decision data of the anomaly detection model that determines that the multimodal feature vector has a violation. This hash value can be a SHA-256 hash value, which is a single data block with a size of 1MB and supports parallel computing, ensuring that the original violation evidence is "fixed as soon as it is generated".

[0160] The smart contract module automatically executes the evidence storage logic based on preset business rules. For example, when a violation such as "expired distributor qualification" is detected, the contract not only triggers the evidence storage transaction, but also synchronously associates the historical evidence storage information of the distribution channel system to form a chain of evidence.

[0161] Consortium blockchain nodes (i.e., blockchain nodes) employ a consensus mechanism (Practical Byzantine Fault Tolerance, PBFT) (consensus latency less than 500ms) to package blocks, and distributed ledger ensures data immutability. Taking telecommunications value-added service orders as an example, after evidence of violations is uploaded to the blockchain, it can be traced through three dimensions: block height, timestamp, and hash value, fully meeting the requirements for judicial evidence in the "Rules for Electronic Data Forensics".

[0162] Figure 13 This is a schematic diagram of the binding of electronic signatures and timestamps provided in an embodiment of the present invention, used to explain the technical implementation and legal validity guarantee logic of electronic signatures and timestamps.

[0163] Here, combined Figure 13 This explains how the technical solution strengthens the legal authority of evidence:

[0164] The electronic signature module uses the national cryptographic SM2 algorithm to realize digital signatures, with a signing time of less than 200ms. It not only meets the requirements of the "Electronic Signature Law" for "reliable electronic signatures", but also uses hardware encryption of private key and supports USBKey or national cryptographic card to prevent signature abuse.

[0165] The timestamp service synchronizes with the National Time Service Center via the Network Time Protocol (NTP) to ensure nanosecond-level accuracy, eliminate the risk of time tampering, and provide an authoritative basis for the "timeliness determination" of evidence.

[0166] The encrypted archive files are protected by AES-256 encryption throughout their entire lifecycle, complying with information security standards. Even if data is leaked during storage or transmission, the original evidence cannot be decrypted.

[0167] Step 104: Obtain the electronic signature and timestamp returned by the blockchain node for the real-time evidence storage information.

[0168] This invention constructs a trusted closed loop for the entire process of "testing, judgment, and evidence storage." By using blockchain for real-time evidence storage, electronic signatures, and timestamps to solidify the validity of evidence, and combining AI preliminary judgment with a dynamic workflow of manual review, it not only solves the problems of easy tampering and difficulty in tracing evidence in existing audits, but also achieves seamless linkage between compliance judgment and trusted evidence storage, providing judicial-grade evidence support for scenarios with high compliance requirements.

[0169] In the AI-based preliminary judgment and manual review workflow, evidence is dynamically updated: the AI ​​preliminary judgment result automatically triggers temporary evidence storage, and after manual review (such as correcting misjudgments or supplementing review opinions), the final conclusion is bound to the electronic signature and timestamp, forming a complete evidence chain of "preliminary judgment-review-final judgment", and each link has an unalterable operation trajectory.

[0170] Through technological integration, multiple values ​​are realized: automating the evidence preservation process improves review efficiency, while blockchain and encryption technologies ensure zero risk of evidence tampering; the time required for manual intervention is reduced from the traditional 4 hours to within 1 hour, and all operations are traceable. The credible evidence system it constructs not only meets the compliance audit requirements of industries such as finance and telecommunications, but also directly supports regulatory investigations and judicial proceedings, significantly enhancing the authority and legal enforceability of compliance testing results.

[0171] It is understood that the distribution channel compliance testing method described in this embodiment of the invention is essentially a dynamic testing method for distribution channel compliance that integrates reinforcement learning. This embodiment also provides a platform applied to this method, namely a dynamic testing platform for distribution channel compliance that integrates reinforcement learning. Through a fully automated process encompassing order snapshot data collection, multimodal data parsing, AI anomaly detection, and automated review and evidence storage, it achieves compliance testing of page elements and processing procedures within the distribution channel system. Furthermore, by integrating AI technology to simulate manual operation, it solves the problems of low efficiency and delayed rule updates in existing manual reviews, covering a complete closed loop from data collection to evidence storage, ensuring compliance and controllability throughout the entire distribution business process.

[0172] Figure 14 This is an overall flowchart of the distribution channel compliance testing method described in this embodiment of the invention, used to illustrate the core steps of the system from order snapshot collection to blockchain evidence storage.

[0173] like Figure 14As shown, the system first acquires order snapshot data, such as screenshots of order pages and order logs, from the distribution channel system via an order snapshot acquisition module. The acquisition frequency supports 100 API calls per second to ensure real-time performance. The acquired order snapshot data is processed by a multimodal parsing engine, which extracts text feature vectors, image feature vectors, and semantic feature vectors using a fusion of OCR, CV, and NLP technologies, generating a 1024-dimensional multimodal feature vector. An AI anomaly detection model uses reinforcement learning algorithms to infer from the multimodal feature vector, calculates compliance scores, and marks anomalies. For example, it triggers a risk warning when it identifies a distribution page's unauthorized use of the term "national-level" in promotional language. The automated review module, combined with a rules engine, generates review conclusions and pushes the violation results to the business system through a link. Simultaneously, it calls the blockchain node's blockchain evidence storage system to perform hash storage of the evidence chain, ensuring data immutability.

[0174] Figure 15 This is a flowchart illustrating the overall process sequence of the distribution channel compliance testing method described in this embodiment of the invention. The flowchart details the data interaction of each module, including the collaborative logic of the order snapshot collection module, the multimodal parsing engine, the AI ​​anomaly detection model, the automated review module, and the blockchain evidence storage system.

[0175] like Figure 15 The interaction details are further refined as follows: The order snapshot collection module requests order snapshot data from the distribution channel system via an interface, generating order snapshot data including screenshots of the order page and order logs. After filtering incremental data within a 5-minute time window, the data is transmitted to the multimodal parsing engine. The multimodal parsing engine transforms unstructured data into structured feature vectors through OCR text recognition, CV image analysis, and NLP semantic understanding, and inputs this data into the AI ​​anomaly detection model. The AI ​​anomaly detection model calculates a compliance score based on a dynamic reward function. If the compliance score is below a threshold, the automated review module is triggered to generate a violation result. Simultaneously, a SHA-256 hash value is generated through the blockchain notarization system and uploaded to the blockchain. Finally, the review result, including an electronic signature and timestamp, is fed back to the distribution channel system. For example, when a telecom package subscription page is found to lack a "first month half price" rule, the system automatically generates a violation evidence package, which is then notarized on the blockchain and pushed to the operations team for review. The entire process takes less than 500ms.

[0176] The overall process of this invention, through a closed-loop design of "collection-parsing-detection-review-storage," achieves full automation of compliance detection for distribution channel systems. In conjunction with the above... Figure 14 and Figure 15It demonstrates the complete logic from data acquisition to intelligent decision-making and evidence consolidation. Among them, the application of multimodal parsing and reinforcement learning improves the detection accuracy in complex scenarios (e.g., text recognition accuracy of 97% and false alarm rate of less than 5%), and blockchain evidence storage ensures the legal validity of compliance conclusions. Compared with the existing manual review, it can greatly improve efficiency, and the timeliness of risk prevention is upgraded from post-event detection to real-time early warning, providing an efficient and reliable technical solution for distribution compliance management in industries such as telecommunications and e-commerce.

[0177] The embodiments of the present invention will be further described below in the following aspects:

[0178] Firstly, in response to the high requirements for transparency and traceability in telecommunications compliance audits, this invention provides an interpretable AI module based on LIME and SHAP, which visualizes the basis for compliance judgments, solves the trust problem of "black box" models, and improves the efficiency of manual review.

[0179] Figure 16 The LIME local interpretation heatmap provided in this embodiment of the invention is used to display the annotation of text violations by the local interpretation heatmap. For example... Figure 16 As shown, the LIME interpreter generates an explanation heatmap for the violating text fragment (such as "medical aesthetics-grade communication service"), marking the keyword "prohibited distribution area" with a weight of 0.35, visually demonstrating the reason for the violation. Reviewers can quickly locate the problem, reducing the review time from 30 minutes to 5 minutes.

[0180] Figure 17 The SHAP value global dependency tree analysis graph provided in this embodiment of the invention is used to analyze feature contribution through dependency tree analysis. For example... Figure 17 As shown, SHAP value calculations reveal that text feature vectors contribute 45% and image feature vectors contribute 30%. Dependency tree visualization exposes the interaction effect of the "limited-time discount icon + first month half-price text". For example, when determining a package page to be in violation, the dependency tree can be used to demonstrate the joint influence of image and text feature vectors, helping the operations team optimize the page design.

[0181] The explainable AI audit decision-making framework combines local explanation with global analysis to transform AI decisions into visually understandable results. Auditor questions can improve location efficiency, reduce false positives, meet the telecommunications industry's regulatory requirements for "explainable and traceable" compliance audits, and enhance human-machine collaboration efficiency.

[0182] Secondly, this invention provides a telecommunications edge-cloud collaborative computing architecture. This architecture is not simply a physical connection between the edge and the cloud, but rather addresses the core pain points of telecommunications distribution channel systems, such as "wide geographical dispersion, large differences in network conditions, and high real-time requirements." It constructs an integrated computing power scheduling system of "intelligent task distribution, dynamic computing power adaptation, and distributed collaboration." Through deep collaboration between edge nodes and the core cloud, it solves the problems of high latency and wasted computing resources in remote areas of existing pure cloud solutions, providing efficient computing power support for compliance testing across all channels.

[0183] Figure 18 This is a schematic diagram of the edge-cloud task scheduling decision tree provided in an embodiment of the present invention, used to clearly present the intelligent decision-making logic of task type and computing power allocation.

[0184] Here, combined Figure 18 This demonstrates that the scheduling logic is deeply adapted to telecommunications service scenarios:

[0185] First, the task types are classified and determined: text parsing (e.g., recognition of package tariff terms), image recognition (e.g., ID card image recognition and qualification document scanning), and model training (e.g., updating a multimodal detection model).

[0186] For lightweight tasks (such as ID card image recognition in offline stores), if the data volume is less than 10MB and the network latency is less than 50ms, it is automatically assigned to edge nodes for execution. Lightweight models (such as OCR engines based on MobileNet architecture) are deployed on the edge nodes to avoid transmission delays when uploading raw data to the cloud.

[0187] For complex tasks (such as multimodal model training and full analysis of violation samples), the process is routed to the core cloud and the powerful computing capabilities of the cloud GPU cluster are utilized to complete the processing.

[0188] For example, in the qualification review of distribution orders for offline stores in remote areas, ID card image recognition is completed locally at the edge node, reducing the processing latency from 800ms in the pure cloud to 250ms, ensuring real-time detection even in scenarios with network fluctuations.

[0189] Figure 19 This is a latency comparison chart of the edge-cloud collaborative architecture provided in the embodiments of the present invention, used to present the intelligent decision-making logic of task type and computing power allocation, and to quantitatively compare the performance differences between edge-cloud collaborative and pure cloud solutions.

[0190] Here, combined Figure 19 The performance leap is clearly demonstrated:

[0191] The edge-cloud collaborative architecture reduces latency by 68.75% compared to pure cloud. In urban and rural channel systems with poor network conditions, the improvement in detection response speed is particularly significant, ensuring the real-time nature of "compliance detection upon order submission".

[0192] By combining CPU / GPU hybrid scheduling strategies (prioritizing CPU for text tasks and automatically scheduling GPU for image tasks), the utilization rate of computing resources can be improved, hardware deployment costs can be reduced, and performance and cost can be perfectly balanced.

[0193] This edge-cloud collaborative architecture precisely adapts to the wide-area coverage characteristics of telecom distribution through a collaborative model of "edge processing of real-time tasks + cloud support for complex computing": edge nodes solve the "last mile" latency problem, while the core cloud ensures large-scale data processing capabilities, and the two achieve seamless collaboration through a unified scheduling protocol. Ultimately, the edge node processing latency is stabilized within 250ms, reducing computing power costs and providing efficient and economical computing power support for compliance testing across the entire telecom system (from provincial agents to township stores), effectively solving the problem of "uneven detection efficiency caused by geographical dispersion".

[0194] Next, the embodiments of the present invention will be described in detail from the following application scenarios:

[0195] Telecom package subscription approval scenario:

[0196] The online ordering pages and processes for various packages offered by telecom operators (such as 5G unlimited packages and family bundled packages) are examined to ensure compliance with package pricing, promotional rules, and user agreements, thereby avoiding issues such as misleading pricing and ambiguous terms and protecting users' right to know and right to choose.

[0197] Figure 20 This flowchart illustrates the compliance detection process for package pricing provided in this embodiment of the invention, showcasing the logic for parsing pricing terms and calculating price discrepancies. Figure 20 As shown, the system crawls the package page every 10 minutes, extracts keywords from the pricing terms (such as "first month half price" or "data rollover") using an NLP module, and calculates the price deviation (with a threshold of ±5%) by comparing it with the compliance rule base. For example, if a package advertises "first month half price" but the actual charge does not reflect this, the system automatically marks it as abnormal and generates an audit conclusion, notifying the operations team to correct the page description in a timely manner.

[0198] Figure 21 This is a schematic diagram illustrating the compliance status of the package ordering process provided in this embodiment of the invention, used to show the status transition from receiving an order to obtaining an audit conclusion. For example... Figure 21As shown, after receiving the order snapshot data, it enters a pending review state. The AI ​​initial judgment module completes the compliance score within 500ms. If the score is greater than 85 points, it automatically passes; otherwise, it proceeds to manual review. The manual review period is 2 hours, and the final review conclusion is output with a chain of evidence. For example, when a user orders a package with a "24-month contract period," the page does not prominently display the cancellation clause, triggering a manual review to avoid potential disputes.

[0199] The embodiments of the present invention improve the compliance rate of tariff labeling from 80% to 98%, reduce the false alarm rate of price deviation to less than 3%, reduce the amount of manual review by 60%, and decrease the user complaint rate by 45%, significantly improving the transparency and compliance of the package ordering process.

[0200] Value-added service compliance audit scenarios:

[0201] For promotional content and subscription processes of telecommunications value-added services (such as data packages, video memberships, or cloud storage), check the compliance of advertising language (e.g., filtering prohibited words) and the integrity of user authorization processes (e.g., secondary confirmation pop-ups) to prevent false advertising and unauthorized activation.

[0202] Figure 22 The word cloud diagram for detecting advertising violations provided in this embodiment of the invention is used to display the distribution of violating keywords. For example... Figure 22 As shown, a violation database of over 5,000 words (such as "national level" and "most favorable") is loaded, and the semantic scanning engine detects the text of value-added service pages in real time with an accuracy rate of 95%. High-frequency violation words (such as "medical aesthetics-grade communication service") are highlighted in red, and the word cloud density reflects the frequency of violations (occurring more than 100 times / day), assisting the operations team in targeted rectification.

[0203] Figure 23 The video playback diagram of the value-added service ordering process provided in this embodiment of the invention is used to present the process of Robotic Process Automation (RPA) simulating user operations and generating evidence packages. For example... Figure 23 As shown, the RPA module simulates user clicks and records the process video at 25fps. Key nodes (such as secondary confirmation pop-ups) are automatically captured as screenshots, generating a three-part evidence chain including video, screenshots, and logs. For example, if a value-added service deducts fees without triggering secondary confirmation, the violation can be located by reviewing the video recording, providing solid evidence for punishment.

[0204] This invention improves the timeliness of advertising violation detection from daily manual inspections to real-time interception, reduces the missed detection rate of prohibited words to less than 2%, increases the compliance rate of the ordering process from 75% to 99%, and reduces the number of complaints about value-added services by 62%, effectively protecting user rights and brand reputation.

[0205] Scenarios for preventing and controlling telecommunications fraud risks:

[0206] In response to high-incidence scenarios of telecommunications network fraud, such as fake prize-winning notifications and orders for counterfeit official applications, the system uses real-time call data monitoring and number correlation analysis to identify abnormal communication behaviors and fraud links, assisting operators in implementing risk blocking.

[0207] Figure 24 The abnormal traffic volume timing detection diagram provided in this embodiment of the invention is used to illustrate the detection of sudden changes in traffic volume, such as... Figure 24 As shown, call data is collected at a frequency of 1 call per minute. An abnormal peak (e.g., a sudden increase of 500 calls per day for a certain number) is identified through a mutation detection algorithm (threshold ±200%), with a detection delay of less than 10 seconds, generating a 24-hour risk curve. For example, if a certain number makes frequent calls to elderly users in the early morning, it is automatically marked as a high-risk target for fraud.

[0208] Figure 25 The fraudulent number association graph provided in this embodiment of the invention is used to present the relationship network between fraudulent numbers and associated numbers. For example... Figure 25 As shown, the knowledge graph module analyzes number communication records to identify the relationship between "caller number - called number - SMS content" (with over 100,000 nodes), and uses community discovery algorithms to locate fraud gangs (e.g., a number communicating intensively with 50 high-risk numbers). Operators can then implement precise shutdowns based on the graph, greatly improving blocking efficiency.

[0209] The embodiments of this invention can achieve an accuracy rate of 91% in identifying fraudulent numbers, shorten the time for intercepting abnormal calls from T+1 days to real time, increase the number of fraud cases solved by public security organs by 28%, and increase the coverage rate of user anti-fraud reminders to 95%, effectively curbing the spread of telecommunications network fraud.

[0210] Automated processing scenario for telecommunications work orders:

[0211] For the entire process of accepting, assigning, and processing service work orders (such as installation, complaints, and fault reports) for telecommunications users, AI is used to analyze work order content, intelligently assign tasks, and automatically verify processing results, thereby improving the efficiency and compliance of work order processing.

[0212] Figure 26 The work order intelligent dispatch rule matrix diagram provided in this embodiment of the invention is used to illustrate the dispatch logic based on work order type, urgency, and region. For example... Figure 26 As shown, based on the work order type (e.g., installation or complaint), urgency level (e.g., high, medium, or low), and regional affiliation (e.g., provincial or municipal), the optimal processing team is matched using Dijkstra's shortest path algorithm. For example, urgent provincial complaint work orders are automatically assigned to provincial customer service experts, reducing processing time from 4 hours to 1 hour.

[0213] Figure 27 This is a histogram comparing work order processing efficiency provided in an embodiment of the invention, used to compare the processing efficiency of manual and AI processes. Figure 27 As shown, traditional manual processing of each order takes an average of 4 hours, while the AI ​​automation solution of this invention reduces the time to 1.5 hours, improving efficiency by 62.5% and increasing the daily processing volume from 500 orders to 1200 orders. Simultaneously, the automatic verification of work order processing results (e.g., installation address compliance) increases the compliance rate from 88% to 98%.

[0214] The embodiments of the present invention reduce the cost of work order processing by 55%, increase user satisfaction from 72% to 89%, and shorten the average time for resolving complaint work orders by 58%, significantly optimizing the quality of telecommunications services and operational efficiency.

[0215] Next, the compatibility design of the application system in the embodiments of the present invention will be explained:

[0216] The system compatibility design addresses the complex IT environment of the telecommunications industry by using a multi-platform interface adaptation layer and browser compatibility technology to achieve seamless compatibility with multiple cloud platforms (such as Huawei Cloud and Tencent Cloud), multiple browsers (such as Chrome and Edge), and heterogeneous terminals, thus solving the problems of high cost and long deployment cycle of cross-platform adaptation of traditional systems.

[0217] Figure 28 This is a multi-platform interface adaptation layer architecture diagram provided in the embodiments of the present invention, used to illustrate protocol conversion and multi-cloud adaptation logic. Figure 29 This is a browser compatibility test coverage diagram provided in an embodiment of the present invention, used to illustrate the browser compatibility test coverage and adaptation technology.

[0218] like Figure 28 and Figure 29 As shown, the protocol conversion module supports RESTful / SOAP protocols and XML / JSON format parsing, enabling heterogeneous interface integration with the telecom's CRM system and Business Operating Support System (BOSS). The multi-cloud adaptation component can encapsulate interface differences between cloud platforms such as Huawei Cloud and Tencent Cloud, for example, calling object storage services from different cloud vendors through a unified interface, avoiding redundant development. The browser adaptation engine uses Polyfill technology to fill in missing interfaces in older browsers, ensuring system functionality consistency in mainstream browsers such as Chrome 80+ and Edge 90+, and the hardware compatibility list covers over 98% of telecom business hall terminal equipment.

[0219] The system compatibility design of this invention, through a layered adaptation architecture and automated testing, achieves cross-platform capability of "develop once, deploy on multiple platforms." Interface adaptation efficiency can be improved by 60% in multi-cloud environments, and the browser compatibility issue resolution cycle is shortened to 2 weeks, significantly reducing the technical integration costs for telecommunications industry customers and supporting their rapid deployment of compliance testing systems in different regions and terminal environments.

[0220] Here, the privacy and security system of this invention is explained:

[0221] The privacy and security system addresses the sensitive nature of telecommunications user data by employing technologies such as data anonymization, access control, and encrypted storage to build full lifecycle security protection. This system meets compliance requirements of the Personal Information Protection Law and the Regulations on the Protection of Personal Information of Telecommunications and Internet Users, preventing data leakage and unauthorized access.

[0222] Figure 30 This is a data desensitization technology roadmap provided for embodiments of the present invention, used to illustrate the sensitive data identification and masking process. For example... Figure 30 As shown, the sensitive field identification module uses regular expressions and NLP techniques (such as named entity recognition) to locate sensitive information such as ID card numbers and mobile phone numbers, achieving an accuracy rate of 99%. The dynamic mask generation module applies an irreversible mask to the middle 4 digits of the mobile phone number and the birthdate field of the ID card number, for example, "1381234". At the same time, it can encrypt and store the desensitized data using the AES-256 algorithm to ensure that even if the data is leaked, the real information cannot be restored.

[0223] Figure 31 The RBAC (Role-Based Access Control) model diagram provided in this embodiment of the invention is used to illustrate role-based access control logic. For example... Figure 31 As shown, user roles are divided into administrators, auditors, and regular users. Administrators have data deletion permissions, auditors can only view and annotate, and regular users can only browse. The permission and resource layers restrict data access through fine-grained controls (e.g., field-level permissions). For example, auditors can only view work order summaries and cannot export complete user information. Role permissions are dynamically bound through a link, meeting the Level 3 requirements of the Cybersecurity Classified Protection System 2.0.

[0224] Therefore, the privacy and security system of this invention combines technical means with management strategies to achieve "usable but invisible" telecommunications user data. The risk of sensitive data leakage can be reduced by 99%, and the interception rate of unauthorized access operations reaches 100%, meeting the stringent requirements of the telecommunications industry for user privacy protection and providing security guarantees for the system's deployment in highly sensitive scenarios.

[0225] Furthermore, the operation, maintenance, and optimization of embodiments of the present invention are explained:

[0226] The operation and maintenance and optimization module uses technologies such as full-link monitoring and incremental model learning to perceive the system's operating status in real time, automatically optimize the detection model and resource allocation, solve the problems of high operation and maintenance complexity and lagging model iteration in traditional AI systems, and ensure the long-term stable and efficient operation of the system.

[0227] Figure 32 This is a schematic diagram of the end-to-end monitoring dashboard provided in this embodiment of the invention, used to display real-time monitoring indicators and early warning logic. The indicator acquisition module collects more than 50 indicators, such as GPU utilization and interface latency, at a frequency of 1 second / time. The early warning engine issues real-time alarms for anomalies such as GPU utilization greater than 85% and latency less than 500ms, with a response time of less than 5 seconds. The monitoring panel displays resource load trends in visual charts. For example, when the CPU utilization of a certain edge node is consistently greater than 90%, the system automatically triggers elastic scaling to ensure that the detection task is not interrupted.

[0228] Figure 33 The incremental learning loop diagram provided in this embodiment of the invention illustrates the iterative process of the model based on new data. New illegal samples (such as emerging fraudulent tactics) are automatically collected daily. High-value data is filtered through an active learning algorithm and injected into the training pipeline for incremental training. For example, when a new "points redemption" fraud pattern emerges in a certain region, the model is updated within 72 hours, increasing the detection accuracy from 70% to 92%, a significant improvement in efficiency compared to manual iteration.

[0229] The operation and maintenance and optimization module of this invention reduces the mean time to repair (MTTR) from 4 hours to 30 minutes and the model lag rate from 15 days to less than 3 days through real-time monitoring and automated iteration. End-to-end monitoring ensures a 35% optimization of resource utilization, and incremental learning improves the system's response speed to emerging violation patterns by 70%, significantly enhancing the system's robustness in complex scenarios in the telecommunications industry.

[0230] Finally, the commercialization model of this invention customized for the telecommunications industry will be described:

[0231] The commercialization model is designed with customized service packages tailored to the characteristics of the telecommunications industry, providing a layered product system of "basic functions + industry plugins + value-added services", supporting both SaaS subscription and local deployment modes, solving the differentiated needs of telecommunications enterprises of different sizes and lowering the technical application threshold.

[0232] Figure 34 This is a functional matrix diagram of a telecommunications industry customized package provided in an embodiment of the present invention, used to illustrate the functional components of the customized package. For example... Figure 34As shown, the basic functionality layer provides general capabilities such as order detection and qualification verification; the telecom-customized layer integrates industry-specific modules such as real-name verification and package rule verification, for example, connecting to the telecom "One-Certificate Verification" system to verify user identity information; the value-added service layer expands scenario-based functions such as fraud prevention and work order automation to meet the diversified needs of operators. Industry rule packages support dynamic loading; for example, the newly released "Administrative Measures for Telecommunications Business Operation Licenses" can be quickly integrated through plugins.

[0233] Figure 35 The cost comparison chart between SaaS subscription and local deployment provided in this embodiment of the invention is used to compare the cost differences between the two deployment modes. Figure 35 As shown, under the SaaS subscription model, the annual cost for fewer than 1,000 channels is less than 50,000 yuan, suitable for small and medium-sized telecom agents. The localized deployment model is for large operators with more than 10,000 channels, with an initial cost starting at 500,000 yuan, but the unit cost decreases significantly as the scale increases. The break-even point is 5,000 channels, and enterprises can choose the optimal solution based on their own scale. For example, a provincial telecom operator adopted localized deployment, covering 20,000 channels, and the annual cost was 40% lower than the SaaS model.

[0234] The commercialization model of this invention, through layered product design and flexible deployment strategies, is adaptable to customers across the entire telecommunications industry chain, from agents to operators. The SaaS model reduces costs for small and medium-sized customers by 70%, while localized deployment meets the customized needs of large enterprises. It is expected to cover more than 90% of telecommunications distribution scenarios, helping operators build intelligent compliance management systems while creating sustainable business value for suppliers.

[0235] In summary, the distribution channel compliance testing method described in this embodiment of the invention includes:

[0236] (1) Dynamic probing capability of multimodal AI fusion:

[0237] By using CV technology to identify page elements (such as compliance logos and product information), NLP to parse text rules (such as contract terms and qualification documents), and RPA to simulate human operations (such as clicks and input), we can achieve full-dimensional compliance inspection of "element detection + process verification + semantic analysis". This breaks through the limitations of traditional single-modal inspection. For example, after extracting handwritten qualification document information through OCR, we can use NLP models to match it with the compliance rule base and automatically mark risk fields.

[0238] (2) Reinforcement learning-driven adaptive testing strategy:

[0239] By introducing reinforcement learning algorithms, test cases are dynamically generated based on the page structure of distribution channels, eliminating the need for manually preset templates. For example, the system can automatically identify newly added buttons on the page and generate click operation tests to verify their compliance (e.g., whether access control complies with rules), solving the problem of traditional testing systems relying on fixed scripts and improving adaptability to dynamic interfaces (e.g., real-time promotional pages, multi-tenant customized interfaces).

[0240] (3) Intelligent scheduling and in-memory computing collaboration of heterogeneous resources:

[0241] Build a CPU / GPU hybrid resource pool and dynamically allocate computing power based on task type (e.g., CPU for text analysis, GPU for image recognition), and combine edge computing to reduce data transmission latency. For example, image comparison tasks are automatically accelerated by GPU, reducing processing time from 500ms to less than 200ms, optimizing computing power allocation efficiency, reducing enterprise deployment costs, and improving real-time detection capabilities.

[0242] (4) Configurable industry-level compliance rule engine:

[0243] The design incorporates a pluggable rules engine that supports the rapid import of compliance rules from multiple industries, including retail, finance, and telecommunications (such as GDPR and e-commerce law). Industry experts can customize rules (such as distribution geographical restrictions and price compliance thresholds) through a visual interface, breaking through the "one-size-fits-all" compliance detection and achieving flexible adaptation across industries and scenarios. For example, the financial industry can configure "dual recording" process verification rules, and the e-commerce industry can configure "false advertising keyword filtering" rules.

[0244] (5) Interpretable architecture for human-machine collaboration:

[0245] By integrating interpretable AI modules such as LIME and SHAP, compliance judgment criteria (e.g., "a certain contract clause violates Article X of XX Regulation") can be visualized. Auditors can quickly locate problems and intervene for review, improving the transparency of AI decision-making, reducing the trust cost of human beings to "black box" models, and ensuring compliance accuracy in complex scenarios (such as the interpretation of ambiguous regulatory clauses).

[0246] Furthermore, the method for testing the compliance of distribution channels according to embodiments of the present invention further includes:

[0247] (1) Multimodal AI detection method:

[0248] A reinforcement learning-based dynamic multimodal detection method: It triggers incremental parsing by listening to DOM changes, combines a knowledge graph in the distribution domain to realize cross-modal business rule verification, and integrates CV, NLP, and RPA to realize a comprehensive technical solution for compliance detection of distribution channel pages and processes, including collaborative logic of element recognition, process simulation, and semantic analysis.

[0249] (2) Dynamic test case generation algorithm:

[0250] The adaptive probing strategy based on reinforcement learning includes state space modeling (e.g., page element states), action space design (e.g., simulated operation types), and reward function definition (e.g., compliance scoring).

[0251] (3) Heterogeneous resource scheduling system:

[0252] The dynamic scheduling mechanism for CPU / GPU hybrid resource pools includes task type identification algorithms, computing power allocation strategies, and edge computing node collaboration methods.

[0253] (4) Scalable rules engine architecture:

[0254] The technical architecture of the plug-in rule management system includes a rule import interface, a visual configuration interface, and a rule conflict detection algorithm.

[0255] (5) Human-machine collaborative workflow:

[0256] The workflow design of "AI initial judgment + manual review" includes an automatic sorting mechanism for abnormal tasks, interactive logic for manual review interface, and optimization algorithm for review result feedback.

[0257] It should be noted that the distribution channel compliance testing method described in this embodiment of the invention has the following advantages:

[0258] By introducing a reinforcement learning-driven test case generation algorithm, test paths can be automatically generated based on changes in page structure (e.g., identifying new buttons and simulating click operations), without manual intervention, and the dynamic adaptation efficiency is improved by more than 80%.

[0259] Real-time process verification is achieved, and compliance is simultaneously detected in the distribution acceptance process (such as qualification review and order processing), avoiding the lag in post-event supervision and shortening the risk prevention time from T+1 days to real time.

[0260] By integrating OCR, NLP, and CV multimodal AI models, a knowledge graph for the distribution domain is constructed. For example, key information from handwritten contracts is extracted using OCR, and NLP semantic analysis is combined with compliance rule bases to automatically mark risk clauses (such as the phrase "prohibited distribution areas"), improving the recognition accuracy from 92% to 97%.

[0261] Breaking through the limitations of "isolated element detection," this technology links page elements with business processes for verification (e.g., verifying the compliance of price tags while simultaneously verifying the completeness of their corresponding order approval processes), covering "process-level compliance" scenarios that are lacking in existing technologies.

[0262] The system features a plug-in rules engine that supports the rapid import of compliance rules from multiple industries, including retail, finance, and telecommunications (such as GDPR, e-commerce law, and financial "dual recording" requirements). Industry experts can customize rules (such as distribution geographic restrictions and price fluctuation thresholds) through a visual interface, improving configuration efficiency by 60%.

[0263] By employing federated learning technology to achieve cross-enterprise data collaboration, companies in different industries can optimize detection models without sharing raw data, thus solving the siloed architecture problem of traditional solutions that rely on "one system per industry".

[0264] A heterogeneous resource intelligent scheduling system was built to dynamically allocate computing power based on task type. CPUs were used for text analysis, GPUs for image recognition, and edge computing nodes for complex model training. The processing time for image comparison tasks was reduced from 500ms to 200ms, and the computing cost was reduced by 35%.

[0265] By optimizing storage and computing collaboration, data transmission across nodes can be reduced. For example, OCR recognition tasks can be executed directly on edge devices to avoid cloud transmission delays, which is suitable for distribution scenarios with limited network conditions (such as offline stores).

[0266] By integrating LIME and SHAP interpretable AI modules, compliance judgment criteria (such as "violation of Article X of XX Regulation") are presented in the form of visual graphs, reducing the time for auditors to locate problems from 30 minutes to 5 minutes.

[0267] The "AI initial assessment + manual review" workflow was designed, which automatically sorts high-risk tasks (such as missing qualification documents or distribution beyond the scope), while manual review is only used for complex scenarios. This improves collaboration efficiency by 50% and reduces the false alarm rate from 15% to below 5%.

[0268] See Figure 36 This invention also provides a device for testing the compliance of distribution channels, comprising:

[0269] The parsing module 3601 is used to parse the order snapshot data within the current time window of the distribution channel system to obtain multimodal feature vectors;

[0270] The detection module 3602 is used to detect whether there are any violations in the distribution channel system based on the multimodal feature vector.

[0271] The evidence storage module 3603 is used to generate a hash value based on the order snapshot data and the violation information of the distribution channel system when a violation is detected in the distribution channel system, and upload the hash value to the blockchain node for evidence storage to form real-time evidence storage information;

[0272] The acquisition module 3604 is used to acquire the electronic signature and timestamp returned by the blockchain node for the real-time evidence storage information.

[0273] Optionally, the distribution channel compliance testing device further includes:

[0274] The data acquisition module is used to retrieve the order list from the customer relationship management system at a preset frequency.

[0275] The first acquisition module is used to obtain multiple initial order snapshot data from the distribution channel system based on the order list;

[0276] The filtering module is used to filter out updated and / or changed order snapshot data from multiple initial order snapshot data within the current time window.

[0277] Optionally, in the distribution channel compliance testing device, the parsing module 3601 includes:

[0278] The first parsing unit is used to parse the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system to obtain text feature vectors.

[0279] The second parsing unit is used to perform image parsing on the compliance identifier in the order snapshot data to obtain image feature vectors;

[0280] The third parsing unit is used to perform semantic parsing on the contract terms in the order snapshot data to obtain semantic feature vectors;

[0281] The concatenation unit is used to concatenate the text feature vector, the image feature vector, and the semantic feature vector to obtain a multimodal feature vector.

[0282] Optionally, in the aforementioned distribution channel compliance testing device, the first parsing unit is specifically used for:

[0283] By monitoring the Document Object Model (DOM) node tree, it can be determined whether the qualification documents and page text in the order snapshot data of the distribution channel system have changed;

[0284] If the qualification document and the page text are determined to have changed, the qualification document and the page text are parsed to obtain text feature vectors.

[0285] Optionally, in the aforementioned distribution channel compliance testing device, the detection module 3602 is specifically used for:

[0286] Based on the multimodal feature vectors and anomaly detection model, detect whether there are any violations in the distribution channel system;

[0287] The anomaly detection model includes a supervised learning module, a semi-supervised learning module, and a reinforcement learning module. The supervised learning module pre-builds a compliance judgment benchmark and uses it to determine whether the multimodal feature vector is in violation based on the compliance judgment benchmark, thereby obtaining a judgment result. The semi-supervised learning module is used to identify the suspected violation type corresponding to the multimodal feature vector, thereby obtaining an identification result. The reinforcement learning module is used to obtain a compliance score for the distribution channel system based on the judgment result and the identification result, and to detect whether the distribution channel system is in violation based on the compliance score.

[0288] Optionally, the distribution channel compliance testing device further includes one of the following:

[0289] The early warning module is used to obtain the risk level of the distribution channel system based on the compliance score, obtain the risk diffusion trend information of the distribution channel system based on the risk level and the risk diffusion trend prediction model, and perform corresponding early warning processing based on the risk diffusion trend information.

[0290] The display module is used to obtain the mapping color of the distribution channel system in the risk warning heat map based on the compliance score, and to display the distribution channel system in the risk warning heat map based on the mapping color.

[0291] Optionally, the distribution channel compliance testing device further includes:

[0292] The transmission module is used to obtain historical violation evidence related to the distribution channel system based on a pre-built violation transmission chain; the violation transmission chain is constructed based on distributor information, qualification documents, and violation types extracted from historical order snapshot data;

[0293] The second acquisition module is used to obtain violation information of the distribution channel system based on the historical violation evidence and the decision data that the multimodal feature vector is in violation, determined by the anomaly detection model.

[0294] Optionally, in the aforementioned distribution channel compliance testing device, the evidence storage module 303 is specifically used for:

[0295] Obtain historical evidence information associated with the aforementioned violation information;

[0296] Triggering a notarization transaction according to preset business rules, the hash value and the historical notarization information are uploaded to the blockchain node for notarization, forming real-time notarization information.

[0297] It should be noted that the apparatus provided in the embodiments of the present invention can implement all the method steps implemented in the above-mentioned product recommended method embodiments and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail.

[0298] This invention also provides a testing device for distribution channel compliance, such as... Figure 37 As shown, it includes:

[0299] The processor 3701, memory 3702, transceiver 3703, and a program or instructions stored in the memory 3702 and executable on the processor 3701; when the processor 3701 executes the program or instructions, it implements the various processes of the above-described distribution channel compliance testing method embodiment and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0300] The transceiver 3703 is used to receive and send data under the control of the processor 3701.

[0301] Among them, Figure 37 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically connecting various circuits of one or more processors represented by processor 3701 and memory represented by memory 3702. The bus architecture can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 3703 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. For different user equipment, the user interface 3704 can also be an interface capable of connecting external or internal devices, including but not limited to keypads, displays, speakers, microphones, joysticks, etc.

[0302] The processor 3701 is responsible for managing the bus architecture and general processing, while the memory 3702 can store the data used by the processor 3701 when performing operations.

[0303] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described distribution channel compliance testing method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0304] This invention also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described distribution channel compliance testing method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0305] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0306] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0307] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for testing the compliance of distribution channels, characterized in that, include: The order snapshot data within the current time window of the distribution channel system is parsed to obtain multimodal feature vectors; Based on the multimodal feature vector, detect whether there are any violations in the distribution channel system; If a violation is detected in the distribution channel system, a hash value is generated based on the order snapshot data and the violation information of the distribution channel system, and the hash value is uploaded to a blockchain node for evidence storage, forming real-time evidence storage information; Obtain the electronic signature and timestamp returned by the blockchain node for the real-time evidence storage information.

2. The method according to claim 1, characterized in that, The method further includes: Retrieve order lists from the customer relationship management system at a preset frequency; Based on the order list, obtain multiple initial order snapshot data from the distribution channel system; Within the current time window, updated and / or changed order snapshot data are filtered from multiple initial order snapshot data.

3. The method according to claim 1, characterized in that, The order snapshot data within the current time window of the distribution channel system is parsed to obtain multimodal feature vectors, including: The text feature vector is obtained by parsing the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system. Image feature vectors are obtained by parsing the compliance identifiers in the order snapshot data; Semantic parsing is performed on the contract terms in the order snapshot data to obtain semantic feature vectors; The text feature vector, the image feature vector, and the semantic feature vector are concatenated to obtain a multimodal feature vector.

4. The method according to claim 3, characterized in that, Text parsing is performed on the qualification documents and page text in the order snapshot data within the current time window of the distribution channel system to obtain text feature vectors, including: By monitoring the Document Object Model (DOM) node tree, it can be determined whether the qualification documents and page text in the order snapshot data of the distribution channel system have changed; If the qualification document and the page text are determined to have changed, the qualification document and the page text are parsed to obtain text feature vectors.

5. The method according to claim 1, characterized in that, Based on the multimodal feature vector, detect whether there are violations in the distribution channel system, including: Based on the multimodal feature vectors and anomaly detection model, detect whether there are any violations in the distribution channel system; The anomaly detection model includes a supervised learning module, a semi-supervised learning module, and a reinforcement learning module. The supervised learning module pre-builds a compliance judgment benchmark and uses it to determine whether the multimodal feature vector is in violation based on the compliance judgment benchmark, thereby obtaining a judgment result. The semi-supervised learning module is used to identify the suspected violation type corresponding to the multimodal feature vector, thereby obtaining an identification result. The reinforcement learning module is used to obtain a compliance score for the distribution channel system based on the judgment result and the identification result, and to detect whether the distribution channel system is in violation based on the compliance score.

6. The method according to claim 5, characterized in that, The method further includes one of the following: The risk level of the distribution channel system is obtained based on the compliance score. Based on the risk level and the risk diffusion trend prediction model, the risk diffusion trend information of the distribution channel system is obtained. Corresponding early warning processing is then performed based on the risk diffusion trend information. The distribution channel system is mapped to a color in the risk warning heatmap based on the compliance score, and the distribution channel system is displayed in the risk warning heatmap based on the mapped color.

7. The method according to claim 5, characterized in that, The method further includes: Based on a pre-constructed violation transmission chain, historical violation evidence related to the distribution channel system is obtained; the violation transmission chain is constructed based on distributor information, qualification documents, and violation types extracted from historical order snapshot data. Based on the historical evidence of violations and the decision data determined by the anomaly detection model that the multimodal feature vector contains violations, violation information of the distribution channel system is obtained.

8. The method according to claim 1, characterized in that, The hash value is uploaded to a blockchain node for evidence storage, forming real-time evidence storage information, including: Obtain historical evidence information associated with the aforementioned violation information; Triggering a notarization transaction according to preset business rules, the hash value and the historical notarization information are uploaded to the blockchain node for notarization, forming real-time notarization information.

9. A device for testing the compliance of distribution channels, characterized in that, include: The parsing module is used to parse the order snapshot data within the current time window of the distribution channel system to obtain multimodal feature vectors; The detection module is used to detect whether there are any violations in the distribution channel system based on the multimodal feature vector. The evidence storage module is used to generate a hash value based on the order snapshot data and the violation information of the distribution channel system when a violation is detected in the distribution channel system, and upload the hash value to the blockchain node for evidence storage to form real-time evidence storage information; The acquisition module is used to acquire the electronic signature and timestamp returned by the blockchain node for the real-time evidence storage information.

10. A testing device for distribution channel compliance, characterized in that, include: A processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the processor executes the program or instructions to implement the distribution channel compliance testing method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for testing the compliance of distribution channels as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the distribution channel compliance testing method as described in any one of claims 1 to 8.

Images