Digital economy-driven artwork online transaction intelligent analysis and management system

Abstract

The invention, which relates to the technical field of artificial intelligence, discloses a digital economy-driven artwork online transaction intelligent analysis and management system comprising a data acquisition module, an event labeling module, a behavior chain construction module, a mode identification module and an early warning module. The data acquisition module is used for acquiring image data, description text data, behavior video data and audio data, the event labeling module is used for identifying and displaying starting and ending nodes, suspicious behavior nodes, speed fluctuation nodes, emotion abnormal nodes and fuzzy description semantic nodes, and the behavior chain construction module is used for constructing and displaying a behavior chain. The mode recognition module is used for generating a matching degree score according to the display behavior chain, and the early warning module is used for generating a counterfeit product sales early warning signal according to the matching degree score; the artwork live sales risk early warning method has the beneficial effects that the real-time performance and accuracy of artwork live sales risk early warning can be improved, the camouflage behavior of an anchor is identified, and the risk that consumers are cheated is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to an online transaction intelligent analysis management system for art driven by digital economy. BACKGROUND

[0002] With the rapid development of digital economy and online live broadcast technology, the transaction mode of art is gradually shifting from traditional offline exhibition to online live broadcast sales. Especially under the promotion of new channels such as short video platforms and social e-commerce, art live broadcast sales has rapidly risen as a new transaction hotspot due to its high transmission efficiency, strong interactivity and wide user reach.

[0003] At present, there are problems such as loose supervision and auditing standards and lack of professional identification power in the art live broadcast sales scene. Most platforms only conduct basic compliance auditing on live broadcast content, lack effective means to screen art authenticity, and ordinary consumers lack basic identification knowledge and often make decisions based on anchor rhetoric or shooting images, resulting in a high risk of being cheated.

[0004] Therefore, an online transaction intelligent analysis management system for art driven by digital economy is proposed. SUMMARY

[0005] In view of the above prior art situation, the present application is proposed. Embodiments of the present application provide an online transaction intelligent analysis management system for art driven by digital economy, which can improve the real-time and accuracy of art live broadcast sales risk early warning, identify anchor disguising behavior, and reduce the risk of consumers being cheated.

[0006] According to one aspect of the present application, an online transaction intelligent analysis management system for art driven by digital economy is provided, comprising: a data acquisition module configured to synchronously acquire image data of an art, description text data, and behavior video data and audio data of an anchor from a target live broadcast stream; an event labeling module configured to identify a display start-stop node of a target art according to the image data, identify a behavior node of the anchor in the live broadcast that matches a preset suspicious action library according to the behavior video data of the anchor, and identify a speech speed fluctuation node, an emotional anomaly node and a fuzzy description semantic node in the commentary process of the anchor according to the audio data and the description text data; a behavior chain construction module configured to construct a display behavior chain generated by the anchor around the art in the live broadcast based on the node recognition result output by the event labeling module; a pattern recognition module configured to match the display behavior chain with a preset fake behavior pattern library to generate a matching degree score; and an early warning module configured to judge whether the matching degree score is higher than a preset score threshold, and if so, generate a potential fake sales early warning signal.

[0007] Compared with existing technologies, the digital economy-driven online art transaction intelligent analysis and management system according to the embodiments of this application can effectively identify abnormal behaviors and spoofing strategies in live broadcasts by collecting and analyzing anchor behavior data and art information in real time, constructing display behavior chains and matching them with counterfeit behavior patterns, and dynamically adjusting warning thresholds in conjunction with anti-spoofing mechanisms. This has the advantages of improving the accuracy of risk warnings and reducing the risk of consumers being deceived. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a block diagram of the intelligent analysis and management system for online art transactions driven by the digital economy, as described in this invention.

[0010] Figure 2 This is a flowchart illustrating the workflow of the digital economy-driven intelligent analysis and management system for online art transactions.

[0011] Figure 3 This is a flowchart illustrating the anti-spoofing module of the intelligent analysis and management system for online art transactions driven by the digital economy, as described in this invention. Detailed Implementation

[0012] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0013] Exemplary System

[0014] Figure 1The illustration depicts a digital economy-driven intelligent analysis and management system for online art transactions according to an embodiment of this application, comprising: a data acquisition module for synchronously acquiring image data and descriptive text data of the artwork, as well as video and audio data of the streamer's behavior from the target live stream; an event annotation module for identifying the start and end points of the artwork's display based on the image data, identifying behavioral nodes performed by the streamer during the live stream that match a preset database of questionable actions based on the streamer's behavioral video data, and identifying nodes of fluctuating speech rate, abnormal emotions, and ambiguous descriptive semantics during the streamer's explanation based on the audio and descriptive text data; a behavior chain construction module for constructing a display behavior chain generated by the streamer around the artwork during the live stream based on the node recognition results output by the event annotation module; a pattern recognition module for matching the display behavior chain with a preset database of counterfeit behavior patterns to generate a matching score; and an early warning module for determining whether the matching score is higher than a preset score threshold, and if so, generating a potential counterfeit sales early warning signal.

[0015] The data acquisition module is a component that acquires images and descriptive text of artworks, as well as video and audio data of the streamer's behavior in real time from the live stream. This can be achieved using video stream capture, speech recognition, and image capture technologies, providing a multimodal data foundation for subsequent analysis. The event annotation module is a component that identifies key nodes in the acquired data. This can be achieved using image recognition algorithms, video action matching models, and natural language processing technologies, used to locate the time range of artwork display and identify abnormal behavior and fluctuations in the streamer's language expression. The behavior chain construction module is a component that integrates the annotated nodes in chronological order. This can be achieved using time-series data modeling techniques, used to generate a structured sequence reflecting the correlation between the streamer's behavior and language expression. The pattern recognition module is a component that matches the behavior chain with a pre-defined abnormal pattern library. This can be achieved using feature vector similarity calculation and pattern matching algorithms, used to quantify the correlation between the current live stream behavior and counterfeit product sales.

[0016] The core innovation of this application lies in constructing a display behavior chain through multi-dimensional data synchronous collection and fusion analysis, and dynamically matching it with a preset abnormal pattern library to achieve automated risk warning in the art live-streaming transaction scenario. The system combines image recognition, action analysis, semantic understanding and pattern matching technologies to form a closed-loop detection mechanism covering the characteristics of the anchor's behavior, language expression and art display, thereby solving the problems of low efficiency and insufficient professional identification ability of manual supervision in the traditional review mechanism.

[0017] The working process and principle of this application are as follows: The data acquisition module synchronously collects image data and descriptive text data of the artwork, as well as video and audio data of the anchor's behavior from the target live stream. This multimodal data is transmitted to the event annotation module for processing. The event annotation module first identifies the start and end nodes of the display of the target artwork based on the image data to determine the display time period of the artwork in the live stream. Then, the module analyzes the anchor's behavioral video data to identify behavioral nodes performed by the anchor in the live stream that match the preset suspicious action library. At the same time, the event annotation module also uses audio data and descriptive text data to identify nodes of fluctuating speech rate, abnormal emotions, and ambiguous descriptive semantics in the anchor's explanation process. These identification results are transmitted to the behavior chain construction module. The behavior chain construction module constructs the display behavior chain generated by the anchor around the artwork in the live stream based on the node identification results output by the event annotation module. This behavior chain contains the complete process of the anchor displaying the artwork and covers information in multiple dimensions such as visual, speech, and text. The completed behavior chain is transmitted to the pattern recognition module. The pattern recognition module matches the display behavior chain with the preset counterfeit behavior pattern library to generate a matching score. This process involves the comparison of behavior patterns and similarity calculation. The generated matching score is then passed to the early warning module, which determines whether the matching score is higher than a preset score threshold. If the matching score exceeds the threshold, the early warning module generates a potential counterfeit sales warning signal. This signal can be used to remind platform administrators or consumers of the potential risk of counterfeit transactions.

[0018] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0019] The data acquisition module uses multi-threaded parallel processing technology to simultaneously acquire image, audio, video, and text data from the live stream. Image data is sampled at 30 frames per second, audio data at 16kHz, video data at 1080p resolution, and text data is acquired through real-time speech-to-text technology.

[0020] The event labeling module uses deep learning algorithms to process the collected data. For image data, it uses object detection algorithms to identify the time points when artworks appear and disappear. For behavioral video data, it uses human pose estimation algorithms to identify the anchor's actions and matches them with a preset database of questionable actions. For audio data, it uses speech emotion analysis algorithms to identify speech rate fluctuations and emotional abnormalities. For text data, it uses natural language processing technology to identify ambiguous descriptive semantics.

[0021] The behavior chain construction module arranges various labeled events in chronological order to build a behavior sequence containing multi-dimensional information. Each sequence element includes a timestamp, event type, and event description.

[0022] The pattern recognition module uses a sequence matching algorithm to compare the constructed behavior chain with the patterns in the counterfeit behavior pattern library. It uses a dynamic time warping algorithm to calculate the similarity between the behavior chain and each pattern, and selects the highest similarity as the matching score.

[0023] The early warning module sets a scoring threshold of 0.8. When the matching score exceeds this threshold, the system generates an early warning signal, which includes the early warning level, the matching score, and a description of the corresponding counterfeit product behavior pattern.

[0024] Through the above-mentioned scheme, this application realizes real-time analysis and risk warning of multimodal data in the process of live-streaming art sales. This multi-level and multi-dimensional analysis method improves the system's accuracy and response speed in identifying counterfeit transactions, and provides a more effective risk management tool for online art trading platforms.

[0025] In some of the solutions mentioned above in this application, the matching score is based solely on the direct matching between the displayed behavior chain and the preset counterfeit behavior pattern. However, the streamer may deliberately adjust the behavior pattern or language expression to avoid abnormal behavior detection, resulting in the score failing to accurately reflect the real risk.

[0026] This application further proposes an anti-spoofing module, which generates a streamer behavior profile before the early warning module determines whether the matching score is higher than a preset score threshold, and adjusts the matching score based on the streamer behavior profile, audio data, image data, behavioral video data, and descriptive text data.

[0027] The anti-masking module includes: a historical profile unit, used to acquire and update the streamer's historical live streaming behavior data and sales records, and to generate a streamer behavior profile based on the historical live streaming behavior data and sales records; a semantic consistency verification unit, used to identify potential conflicts between the streamer's semantics and the visual information of the target artwork based on audio data and image data, and to obtain a consistency index; an anti-masking model unit, used to generate a probability of masking that represents the streamer's deliberate avoidance of abnormal behavior based on behavioral video data, audio data, and descriptive text data through a pre-trained adversarial model; and an enhanced early warning unit, used to generate a scoring adjustment factor based on the streamer's behavior profile, consistency index, and masking probability through a pre-trained decision model, and to adjust the matching score based on the scoring adjustment factor and a preset adjustment function.

[0028] Specifically, the historical profiling unit normalizes data such as the frequency and duration of suspicious actions in the streamer's historical live-streaming behavior data, as well as the return rate in sales records. It then generates a fixed-dimensional profile using a machine learning encoding model to characterize the streamer's long-term behavioral features. The semantic consistency verification unit calculates the similarity between textual semantic keywords obtained from speech recognition and visual semantic tags obtained from image recognition. If semantic conflicts are significant, the consistency index decreases. The anti-disguise model unit analyzes the streamer's micro-expressions in the video, tone fluctuations in the audio, and logical loopholes in the descriptive text using an adversarial model, outputting a disguise probability. A higher probability indicates a greater likelihood that the streamer is deliberately concealing abnormal behavior. The enhanced early warning unit inputs the profile, consistency index, and disguise probability into the decision model to generate a scoring adjustment factor. For example, when the disguise probability exceeds a preset threshold, the adjustment factor increases. Finally, the original score is combined with the adjustment factor using a formula to obtain a more accurate matching score. Thus, the system can effectively identify streamer disguising behavior, reduce the risk of misjudgment, and improve the reliability of early warning signals.

[0029] In some of the solutions described above in this application, generating a streamer behavior profile based on historical live streaming behavior data and sales records specifically includes: extracting behavioral feature data containing the frequency and duration of suspicious actions from historical live streaming behavior data and a preset suspicious action library; extracting sales feature data containing sales success rate, return rate, and number of high-risk transactions from historical sales records; normalizing the behavioral feature data and sales feature data to generate a unified numerical feature set; and generating a fixed-dimensional streamer behavior profile based on the numerical feature set using a preset machine learning encoding model.

[0030] Among them, behavioral feature data is multidimensional time series data generated by statistically analyzing the number of times each action in the preset suspicious action library appears in historical live broadcasts and the duration of each occurrence; sales feature data is discrete numerical data generated by calculating the percentage of successful transactions, the percentage of returned orders, and the number of transactions marked as high risk in historical transaction records; normalization processing adopts the maximum-minimum scaling method to linearly transform data of different dimensions to the [0,1] interval; the machine learning encoding model adopts an autoencoder structure, learns the low-dimensional representation of the numerical feature set through training data, and generates a fixed-length vector as the anchor's behavioral profile.

[0031] Specifically, the frequency and duration of questionable actions are statistically analyzed using sliding time windows. Each window records the number of times a preset action is triggered and its cumulative duration, forming a composite feature matrix of time and action dimensions. Sales success rate is calculated as the ratio of successful orders to total orders, return rate as the ratio of returned orders to successful orders, and the number of high-risk transactions is directly obtained from the number of orders marked by the platform's risk control system. During normalization, the maximum and minimum values ​​of each feature dimension are pre-calculated and stored from the training set data, and a linear transformation is applied to the input data during real-time processing. In the training phase, the autoencoder model learns the latent spatial representation of the numerical feature set by minimizing the error between the input and the reconstructed output. In the inference phase, the encoder maps the normalized numerical feature set into a dense vector of fixed dimensions. This vector simultaneously preserves the key statistical regularities and distribution characteristics of the original data, providing stable and computable feature input for subsequent score adjustments.

[0032] Through the above technical solutions, this application achieves systematic analysis and quantitative representation of the anchor's historical behavior and sales data. The resulting anchor behavior profile can comprehensively reflect the anchor's behavioral patterns and sales characteristics, providing a reliable data foundation for subsequent risk assessment and early warning. Furthermore, generating fixed-dimensional behavior profiles through machine learning encoding models improves the comparability of data and the universality of the model, which is beneficial for large-scale application and cross-platform analysis.

[0033] In some of the solutions described above in this application, identifying potential conflicts between the semantic information of the broadcaster and the visual information of the target artwork based on audio data and image data specifically includes: performing speech recognition on the audio data to obtain text semantic information; extracting keywords and labeling semantic information on the text semantic information to generate a text semantic feature set; performing image recognition on the image data to obtain visual feature information of the target artwork; mapping the visual feature information to corresponding semantic labels to generate a visual semantic feature set; and calculating the semantic similarity between the text semantic feature set and the visual semantic feature set to obtain a consistency index.

[0034] The process involves several key steps: Speech recognition uses automatic speech recognition technology to convert audio into text; keyword extraction uses natural language processing algorithms to filter terms related to the artwork's attributes, such as material, era, and artist's name; semantic tagging classifies keywords based on a pre-defined tagging system, forming a text semantic feature set containing multi-dimensional attributes; image recognition uses convolutional neural networks to extract visual features of the artwork, such as color distribution, brushstroke texture, and shape contours; visual feature information is mapped to structured semantic descriptions corresponding to text semantic tags through a pre-trained classification model, for example, mapping oil painting brushstroke features to the tag "oil painting technique"; and semantic similarity calculation uses a cosine similarity algorithm to compare the vectorized representations of the text semantic feature set and the visual semantic feature set, outputting a consistency index within the range of 0 to 1.

[0035] For example, when a live streamer describes an artwork as a "19th-century European oil painting," the semantic information generated by speech recognition is extracted to include keywords such as "19th century," "Europe," and "oil painting," and labeled with three semantic tags: era, region, and technique. Simultaneously, image recognition detects features such as the thickness of the paint layer and the direction of brushstrokes in the artwork, which the classification model maps to the tag "watercolor technique." By calculating the cosine similarity between the text semantic feature set and the visual semantic feature set, a conflict is found between the "oil painting" and "watercolor technique" tags, at which point the consistency index significantly decreases. This index, along with the probability of deception and historical behavior profiles, is input into the decision model, triggering dynamic adjustments to the matching score, thereby enhancing the ability to identify behaviors that deliberately conceal the true attributes of artworks.

[0036] In some of the schemes described above in this application, the preset adjustment function is specifically as follows: ;in, The adjusted match score, Score the original match. As a rating adjustment factor, The preset weighting coefficient for controlling the adjustment range is set according to the actual scenario, with a value range of 0 to 1. For example, when set to 0.3, the adjustment range is limited to within 30% of the original score to avoid excessive correction deviation.

[0037] Specifically, the adjustment function merges the score adjustment factor with the original score through linear superposition, achieving a dynamic balance between the two through weighting coefficients. When the score adjustment factor increases, the adjusted score increases proportionally, enhancing the sensitivity of abnormal behavior warnings; when the weighting coefficient decreases, the adjustment range is compressed, avoiding the risk of misjudgment. For example, when high-frequency suspicious actions are detected in the streamer's historical behavior profile, the score adjustment factor is calculated to be 0.5. If the weighting coefficient is set to 0.4, the adjusted score is 1.2 times the original score. This allows for the addition of the comprehensive impact of abnormal behavior while retaining the original matching information, thereby improving the accuracy of the warning signal generation.

[0038] In some of the above-mentioned schemes of this application, constructing the display behavior chain specifically includes: sorting each behavior node by time within the time interval corresponding to the start and end nodes of the display to generate a broadcaster behavior sequence; aligning each speech rate fluctuation node, emotional abnormality node, and ambiguous description semantic node by time within the time interval corresponding to the start and end nodes of the display to generate a language expression sequence; and integrating the broadcaster behavior sequence and the language expression sequence in chronological order to obtain the display behavior chain.

[0039] The time sorting process extracts the timestamps of the behavior nodes and arranges them in chronological order to generate a sequence, ensuring the continuity of the behavior logic. The time alignment process matches the timestamps of the language expression nodes with the time intervals of the display start and end nodes, filtering out invalid nodes that exceed the time interval. The integration process uses time axis synchronization technology to map the behavior sequence and the language expression sequence to the same time dimension, generating a complete behavior chain containing behavior and language interaction information.

[0040] Specifically, the data acquisition module acquires the start and end nodes of the display, defining the time range for the artwork display. The event annotation module identifies both the broadcaster's behavioral nodes and verbal expression nodes, which are then timestamped. Time sorting arranges the behavioral nodes by timestamps, forming a behavioral sequence reflecting the broadcaster's actions. Time alignment confines the verbal expression nodes to the display time range, avoiding interference from non-display periods. During integration, the behavioral and verbal expression sequences are merged through timeline synchronization. For example, a broadcaster's action at a specific moment is associated with a corresponding speech rate fluctuation node, forming a multi-dimensional behavioral chain. Thus, the display behavioral chain comprehensively records the broadcaster's behavior and verbal dynamics during the display, providing a high-precision time-series data foundation for subsequent pattern matching and effectively improving the accuracy of risk warnings.

[0041] In some of the above-mentioned solutions in this application, generating a matching score specifically includes: encoding the anchor behavior sequence and language expression sequence in the displayed behavior chain using temporal features to generate a behavior chain feature vector in a unified format; calling a set of counterfeit behavior templates that match the target artwork from the counterfeit behavior pattern library, and extracting the behavior feature vectors corresponding to each template in the counterfeit behavior template set; calculating the similarity between the behavior chain feature vector and each behavior feature vector according to a similarity measurement algorithm to obtain a similarity set; and calculating the matching score by weighting the top N similarities with the highest similarity in the similarity set, where N is a preset positive integer.

[0042] The temporal feature encoding can employ Long Short-Term Memory (LSTM) networks or Temporal Convolutional Networks (TCNs) to extract dynamic change patterns within the sequence; the set of counterfeit behavior templates is dynamically filtered based on the art category, for example, paintings correspond to templates containing features such as high-frequency hand gestures and contradictory language; the similarity measurement algorithm uses cosine similarity or dynamic time warping to measure the degree of matching between the behavior chain and the template in the time dimension; the selection range for the first N similarities in the weighted calculation can be set to 3 to 5, and the weight coefficient α is optimized to 0.6 to 0.8 based on historical data; the weighted calculation is as follows:

[0043] ;

[0044] in, These are weighting coefficients, with values ​​ranging from [0, 1]. This represents the set of the N highest similarity values ​​in the similarity set. Represents the set of the N highest similarity values. There are several similarity values, and N is set to a range of 3 to 5.

[0045] Specifically, the anchor's behavioral sequence is temporally encoded into a vector containing action frequency and duration, while the language expression sequence is encoded into a vector containing speech rate changes and emotional fluctuations. These two are concatenated to form a behavioral chain feature vector. Each template in the counterfeit behavior template set generates a standardized feature vector using the same encoding method. Through similarity calculation, multiple templates closest to the current behavioral chain are selected, and a weighted average similarity is calculated, combining the highest and average similarities. For example, when α is 0.7, the scoring result highlights both the most significant abnormal patterns and common features across multiple templates. This process, through multi-dimensional feature fusion and dynamic weighting mechanisms, enhances the robustness of identifying deliberately disguised behavior while reducing the risk of misjudgment based on a single similarity index.

[0046] In some of the solutions described above in this application, when the similarity between the displayed behavior chain and a certain counterfeit behavior template is significantly higher than that of other templates, averaging the results will dilute the saliency of the template, reduce the discrimination accuracy, and may mask the identification of high-risk behavior chains.

[0047] This application further proposes that before calculating the matching score, the following steps are included: determining whether the difference between the highest similarity and the second highest similarity in the similarity set is greater than a preset difference threshold. If so, the highest similarity is used as the matching score and the matching score calculation is skipped; otherwise, the matching score calculation continues.

[0048] The preset difference threshold is determined by the feature difference distribution between the counterfeit behavior pattern and the normal behavior pattern in historical data, and the value range is 0.1 to 0.3. The difference judgment process uses a sliding window algorithm to traverse the similarity set and identify the highest and second highest similarity values. When the difference between the highest similarity and the second highest similarity exceeds the threshold, the highest similarity is directly output as the final score; otherwise, the weighted calculation function is called to generate a score.

[0049] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0050] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A digital economy-driven intelligent analysis and management system for online transactions of art pieces, characterized in that, Comprise: a data collection module for synchronously collecting image data of an artwork, description text data, and behavior video data and audio data of an anchor from a target live stream; an event labeling module for identifying display start and end nodes of a target artwork according to the image data, identifying behavior nodes of the anchor in the live stream that match a preset suspicious action library according to the behavior video data of the anchor, and identifying a speech speed fluctuation node, an emotional anomaly node, and a fuzzy description semantic node in the commentary process of the anchor according to the audio data and the description text data; a behavior chain construction module for constructing a display behavior chain of the anchor around the artwork in the live stream based on the node recognition results output by the event labeling module; a pattern recognition module for matching the display behavior chain with a preset fake behavior pattern library to generate a matching degree score; a warning module for judging whether the matching degree score is higher than a preset score threshold, and if so, generating a potential fake sales warning signal.

2. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 1, characterized in that, Further comprise: an anti-camouflage module for generating an anchor behavior portrait before the warning module judges whether the matching degree score is higher than the preset score threshold, and adjusting the matching degree score according to the anchor behavior portrait, audio data, image data, behavior video data, and description text data. 3.The digital economy driven art online transaction intelligent analysis management system according to claim 2, characterized in that, The anti-camouflage module comprises: a historical portrait unit for obtaining and updating historical live behavior data and sales records of the anchor, and generating an anchor behavior portrait according to the historical live behavior data and sales records; a semantic consistency verification unit for identifying a potential conflict between the anchor's semantics and the visual information of the target artwork according to the audio data and the image data, and obtaining a consistency index; an anti-camouflage model unit for generating a camouflage probability representing the anchor's intentional avoidance of abnormal behavior according to the behavior video data, audio data, and description text data through a pre-trained adversarial model; an enhanced warning unit for generating a score adjustment factor according to the anchor behavior portrait, consistency index, and camouflage probability through a pre-trained decision model, and adjusting the matching degree score according to the score adjustment factor and a preset adjustment function.

4. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 3, characterized in that, The generation of the anchor behavior portrait according to the historical live behavior data and sales records comprises: extracting behavior feature data containing the frequency and duration of suspicious action occurrence according to the historical live behavior data and the preset suspicious action library; extracting sales feature data containing the sales success rate, return rate, and number of high-risk transactions according to the historical sales records; normalizing the behavior feature data and the sales feature data to generate a unified numerical feature set; generating a fixed-dimensional anchor behavior portrait according to the numerical feature set through a preset machine learning coding model.

5. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 3, characterized in that, The identification of a potential conflict between the anchor's semantics and the visual information of the target artwork according to the audio data and the image data comprises: performing speech recognition on the audio data to obtain text semantic information; performing keyword extraction and semantic label labeling on the text semantic information to generate a text semantic feature set; performing image recognition on the image data to obtain visual feature information of the target artwork; Map the visual feature information to corresponding semantic labels to generate a visual semantic feature set; Calculate the semantic similarity between the text semantic feature set and the visual semantic feature set to obtain the consistency index.

6. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 3, characterized in that, The preset adjustment function is: ； wherein, is the adjusted matching score, is the original matching score, is the score adjustment factor, is a preset weight coefficient for controlling the adjustment range.

7. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 1, characterized in that, The construction of the display behavior chain includes: Time sequence the behavior nodes within the time interval corresponding to the display start and end nodes to generate a host behavior sequence; Time align the speech speed fluctuation nodes, emotional anomaly nodes, and ambiguous description semantic nodes within the time interval corresponding to the display start and end nodes to generate a language expression sequence; Integrate the host behavior sequence and the language expression sequence in chronological order to obtain the display behavior chain.

8. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 7, characterized in that, The generation of the matching degree score includes: Time sequence feature encode the host behavior sequence and the language expression sequence in the display behavior chain to generate a uniform format behavior chain feature vector; Call a set of counterfeit behavior templates matching the target artwork from the counterfeit behavior pattern library, and extract the behavior feature vector corresponding to each template in the set of counterfeit behavior templates; Calculate the similarity between the behavior chain feature vector and each behavior feature vector according to a similarity measurement algorithm to obtain a similarity set; Calculate the matching degree score according to the top N similarities in the similarity set, where N is a preset positive integer. 9.The digital economy driven intelligent analysis and management system for online transaction of art pieces according to claim 8, wherein, The weighted calculation is specifically: ; wherein, is a weighting coefficient, and the value range is [0, 1], denotes a set of the first N highest similarity values in the set of similarity values, denotes the i-th similarity value in the set of the first N highest similarity values. denotes the i-th similarity value in the set of the first N highest similarity values.

10. The digital economy-driven intelligent analysis and management system for online transactions of artworks according to claim 8 or 9, characterized in that, Before the calculation of the matching degree score, it further includes: Determine whether the difference between the highest similarity and the second highest similarity in the similarity set is greater than a preset difference threshold, if yes, take the highest similarity as the matching degree score and skip the matching degree score calculation, otherwise, continue the matching degree score calculation.

Images